Paper deep dive
Combining General and Domain-Specific Pretext Tasks for Brain MR Image Segmentation
Tasneem Nasser, Susanne Schmid, Roberto Souza, Naser El-Sheimy
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/28/2026, 3:24:59 AM
Summary
This study proposes a multitask self-supervised pretraining framework for brain MR image segmentation that combines a domain-specific pretext task (voxel-level brain age prediction) with a general pretext task (image inpainting). Using SwinUNETR as the backbone, the framework is evaluated on three downstream segmentation tasks: multiple sclerosis lesion segmentation, ischemic stroke lesion segmentation, and cortical brain structure segmentation. Results indicate that the multitask approach consistently outperforms single-task pretraining and training from scratch, demonstrating the benefit of combining domain-specific and general representations for developing generalizable neuroimaging foundation models.
Entities (13)
Relation Signals (11)
Multitask self-supervised pretraining → combines → Voxel-level brain age prediction
confidence 95% · propose a multitask self-supervised pretraining framework that jointly optimizes both objectives [image inpainting and voxel-level brain age prediction]
Multitask self-supervised pretraining → combines → Image inpainting
confidence 95% · propose a multitask self-supervised pretraining framework that jointly optimizes both objectives [image inpainting and voxel-level brain age prediction]
Multitask self-supervised pretraining → evaluatedon → Ischemic stroke lesion segmentation
confidence 95% · ischemic stroke lesion segmentation, and cortical brain structure segmentation
Multitask self-supervised pretraining → evaluatedon → Cortical brain structure segmentation
confidence 95% · ischemic stroke lesion segmentation, and cortical brain structure segmentation
Multitask self-supervised pretraining → evaluatedon → Multiple sclerosis lesion segmentation
confidence 95% · The pretrained models are evaluated on three downstream magnetic resonance image segmentation tasks: multiple sclerosis lesion segmentation
SwinUNETR → usedfor → Multitask self-supervised pretraining
confidence 95% · SwinUNETR was selected as the backbone for the proposed framework.
Multitask self-supervised pretraining → outperforms → Single-task pretraining
confidence 90% · proposed multitask pretraining framework consistently outperformed the single-task pretrained models
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:A key challenge in medical image analysis is the scarcity of large annotated datasets for specific populations and diseases. As deep learning models rely heavily on labeled data, effective transfer learning strategies are needed to reduce the dependence on manual annotations. Self-supervised learning has emerged as a promising approach for developing foundation models by enabling the learning of transferable feature representations from large-scale unlabeled medical imaging datasets. In this study, we investigate voxel-level brain age prediction as a domain-specific self-supervised pretext task and compare it with image inpainting, a widely used non-domain-specific alternative. We further propose a multitask self-supervised pretraining framework that jointly optimizes both objectives to learn complementary neuroimaging representations. The pretrained models are evaluated on three downstream magnetic resonance image segmentation tasks: multiple sclerosis lesion segmentation, ischemic stroke lesion segmentation, and cortical brain structure segmentation. Overall, the proposed multitask pretraining framework consistently outperformed the single-task pretrained models and training from scratch across most experimental settings, demonstrating the benefit of combining domain-specific and general self-supervised learning pretext tasks for the development of generalizable neuroimaging foundation models.\ Code Availability: The source code used in this study is publicly available at this https URL
Tags
Links
- Source: https://arxiv.org/abs/2609.30708v1
- Canonical: https://arxiv.org/abs/2609.30708v1
Trouble viewing inline? Open PDF directly →
Full Text
79,295 characters extracted from source content.
Expand or collapse full text
Combining General and Domain-Specific Pretext Tasks for Brain MR Image Segmentation Tasneem Nasser a,b,∗ , Susanne Schmid b,c , Roberto Souza b,c and Naser El-Sheimy d a Biomedical Engineering University of Calgary Calgary AB Canada b Hotchkiss Brain Institute University of Calgary Calgary AB Canada c Electrical and Software Engineering University of Calgary Calgary AB Canada d Geomatics Engineering University of Calgary Calgary AB Canada A R T I C L E I N F O Keywords: Foundation Models Multiple Sclerosis Stroke Segmentation SwinUNETR Transfer Learning Voxel-Level Brain Age Prediction A B S T R A C T A key challenge in medical image analysis is the scarcity of large annotated datasets for specific populations and diseases. As deep learning models rely heavily on labeled data, effective transfer learning strategies are needed to reduce the dependence on manual annotations. Self-supervised learn- ing has emerged as a promising approach for developing foundation models by enabling the learning of transferable feature representations from large-scale unlabeled medical imaging datasets. In this study, we investigate voxel-level brain age prediction as a domain-specific self-supervised pretext task and compare it with image inpainting, a widely used non-domain-specific alternative. We further propose a multitask self-supervised pretraining framework that jointly optimizes both objectives to learn complementary neuroimaging representations. The pretrained models are evaluated on three downstream magnetic resonance image segmentation tasks: multiple sclerosis lesion segmentation, ischemic stroke lesion segmentation, and cortical brain structure segmentation. Overall, the proposed multitask pretraining framework consistently outperformed the single-task pretrained models and training from scratch across most experimental settings, demonstrating the benefit of combining domain-specific and general self-supervised learning pretext tasks for the development of generalizable neuroimaging foundation models. Code Availability: The source code used in this study is publicly available at GitHub repository. 1. Introduction Data scarcity remains a major challenge in training deep learning (DL) models for medical image segmentation, where labelled data is limited due to the time-consuming na- ture of image annotation and the need for expert knowledge. Since most DL frameworks rely on supervised learning, they require large amounts of annotated data to learn appropriate representations. Developing medical imaging foundation models (FMs) that learn transferable representations from diverse unla- beled medical imaging datasets has emerged as a promising solution to this challenge. Once pretrained, these models can be efficiently adapted to multiple downstream tasks using only limited annotated data [He et al., 2025]. Cross- domain generalization, which refers to the model’s ability to perform well on new tasks or datasets that differ from the data it was initially trained on, for example, adapting a model trained on healthy brain scans to segment tumours or detect neurological disorders, is a key advantage of such models. A FM with this capability can be fine-tuned for downstream tasks using limited annotated data [He et al., 2021]. However, the success of such models relies on the ability to learn robust and transferable representations. Given the difficulty of manual annotation, self-supervised learning has emerged as an alternative paradigm for feature ∗ Corresponding author tasneem.nasser@ucalgary.ca (T. Nasser); susanne.schmid@ucalgary.ca (S. Schmid); roberto.souza2@ucalgary.ca (R. Souza); elsheimy@ucalgary.ca (N. El-Sheimy) ORCID(s): learning [Zheng, 2024]. Self-supervised learning uses pre- text tasks such as learning spatial relationships, predicting image rotations, and masking and then reconstructing image patches to extract useful representations from unlabeled data. Masked Autoencoder pretrains Vision Transformers by randomly masking image regions and predicting the missing content, a technique known as masked image modelling [Kim et al., 2025]. Kim et al. [Kim et al., 2025] extended this concept to Swin Transformers and demonstrated their effectiveness in learning visual representations without re- quiring large labelled datasets. Despite these advances, most existing self-supervised pretext tasks are general (i.e., domain-agnostic), aiming to learn general visual representations rather than anatomi- cally meaningful neuroimaging features. Consequently, their direct application to 3D brain magnetic resonance (MR) images may not fully exploit domain-specific anatomi- cal knowledge. Designing domain-specific pretext tasks is therefore essential to encourage models to learn rep- resentations that are more relevant to medical imaging applications [Kim et al., 2025]. Motivated by this, we previously proposed voxel-level brain age prediction as a domain-specific self-supervised pretext task [Nasser et al., 2026]. Unlike conventional self-supervised pretext tasks that primarily exploit image statistics, voxel-level brain age prediction incorporates domain knowledge by using the assumption that the chronological age is equal to the brain age in healthy subjects as supervision. This encourages the model to learn representations that are anatomically and biologically meaningful, capturing structural variations T. Nasser et al.: Preprint submitted to ElsevierPage 1 of 16 arXiv:2609.30708v1 [cs.CV] 25 Sep 2026 Investigating Pretext Tasks for Brain MR Segmentation associated with the normal aging process. Since chrono- logical age is routinely available as demographic metadata in normative neuroimaging datasets, it provides a practical and annotation-free source of supervision for large-scale neuroimaging pretraining. Therefore, in this work, we aim to take a step further beyond our previous study [Nasser et al., 2026], which investigated voxel-level brain age prediction as a domain- specific pretext task for brain segmentation in data-scarce scenarios. Rather than investigating domain-specific and general pretraining tasks independently, we evaluate both single-task pretraining strategies and a multitask pretraining framework that jointly learns image inpainting and voxel- level brain age prediction. The pretrained models are evaluated on three down- stream MR image segmentation tasks. The first two tasks, multiple sclerosis (MS) lesion segmentation and ischemic stroke lesion segmentation, are not directly associated with brain aging. The last task is cortical brain structure segmen- tation, an age-related task. Evaluating these diverse down- stream tasks allows us to investigate whether voxel-level brain age prediction learns, in addition to a general pretext task (i.e., image inpainting), transferable representations that generalize beyond age-related applications. The main contribution of this work is to demonstrate that voxel-level brain age prediction, particularly when com- bined with multitask pretraining, serves as an effective self- supervised pretext task for brain MR image segmentation across both age-related and non-age-related downstream applications. 2. Methodology 2.1. Proposed Framework This work investigates combining voxel-level brain age prediction, a domain-specific self-supervised pretext task, with image inpainting, a general pretext task, to build a brain MR foundation model. Based on preliminary architecture- selection experiments conducted using the voxel-level brain age prediction task, SwinUNETR was selected as the back- bone for the proposed framework. We compare this com- bined approach with models pretrained using image inpaint- ing alone and voxel-level brain age prediction alone. We evaluate the pretrained models on three downstream MR image segmentation tasks. MS lesion segmentation and is- chemic stroke lesion segmentation serve as non-age-related tasks, while cortical brain structure segmentation serves as an age-related task. This choice was deliberate, since the voxel-level brain age prediction pretext may not general- ize well to non-age-related downstream tasks. Because the downstream tasks require different MR input sequences, each pretraining strategy is designed to match the sequences required by its corresponding downstream task. Figure 1 presents an overview of the proposed framework. 2.2. Datasets 2.2.1. Datasets for Pretext Tasks Two pretraining dataset configurations were selected according to the imaging modalities required by the cor- responding downstream segmentation tasks. Paired T1- weighted (T1w) and Fluid-Attenuated Inversion Recovery (FLAIR) MR images were used for pretraining prior to MS lesion segmentation, whereas T1w-only images were used for pretraining prior to cortical brain structure and ischemic stroke lesion segmentation. A summary of the pre- training datasets, including their age and sex distributions, is presented in Figure 2. Each dataset configuration was used to train the three self-supervised pretraining strategies: voxel-level brain age prediction, image inpainting, and the proposed multitask framework, before fine-tuning on the corresponding downstream segmentation tasks. Pretraining dataset for MS lesion segmentation Since MS lesion segmentation requires both T1w and FLAIR MR scans, the three self-supervised pretraining strategies were trained using paired T1w and FLAIR scans from the Calgary Normative Study (CNS) [McCreary et al., 2020] (Figure 2b). The dataset was divided into 423 training, 122 validation, and 63 test scans for self-supervised pretrain- ing. Pretraining dataset for ischemic stroke lesion seg- mentation and brain cortical segmentation Cortical brain structure segmentation and ischemic stroke lesion segmentation require only T1w MR scans. There- fore, we compiled a multi-source dataset of normative T1w MR scans from individuals aged 18–80 years using sev- eral publicly available datasets, as shown in Figure 2a. These included CORR [Zuo et al., 2014], IXI [IXI Dataset], ABIDE I [Di Martino et al., 2014], ABIDE I [Di Martino et al., 2017], OASIS-1 [Marcus et al., 2007], ADNI [Jack et al., 2008], Cam-CAN [Shafto et al., 2014], the Calgary- Campinas dataset [Souza et al., 2018], and the CNS [Mc- Creary et al., 2020]. To encourage the pretrained models to learn representa- tive features across the adult lifespan, the dataset was curated to maintain a balanced age and sex distribution across the training, validation, and test sets. This balanced sampling re- duces demographic bias during pretraining and promotes the learning of more generalizable feature representations. The resulting multi-source T1w pretraining dataset comprised 1,710 MR scans, including 1,025 training, 428 validation, and 257 test scans. 2.2.2. Dataset for MS Lesion Segmentation The MS lesion segmentation experiments were per- formed using the MSLesSeg dataset [Guarnera et al., 2025]. The dataset comprises 115 MR scans from 75 patients with MS and includes paired T1w and FLAIR images together with expert manual lesion annotations. T. Nasser et al.: Preprint submitted to ElsevierPage 2 of 16 Investigating Pretext Tasks for Brain MR Segmentation Figure 1: Overview of the proposed framework. The figure illustrates the self-supervised pretraining strategies and their transfer to downstream MR segmentation tasks. Three pretraining strategies are investigated: (1) single-task image inpainting, in which masked MR volumes are used as input and the model reconstructs the original image; (2) single-task voxel-level brain age prediction, in which the full MR volume is used to predict a 3D voxel-wise brain age map; and (3) multitask pretraining, which jointly learns both image inpainting and voxel-level brain age prediction, producing both reconstructed MR and voxel-wise age map outputs. The pretrained SwinUNETR model is subsequently fine-tuned on three downstream segmentation tasks: multiple sclerosis lesion segmentation, and stroke lesion segmentation (non age-related), cortical brain structure segmentation (age-related) (a)(b) Figure 2: Age and sex distributions of the pretraining datasets. (a) Multi-source T1w dataset used for pretraining prior to ischemic stroke lesion and cortical brain structure segmentation. (b) Calgary Normative Study (CNS) dataset, containing paired T1w and FLAIR images, used for pretraining prior to MS lesion segmentation. 2.2.3. Dataset for Ischemic Stroke Lesion Segmentation We utilized the ISLES 2026 Challenge training dataset [noa, Absher et al., 2024], which comprises 1,453 cases collected from multiple centers. To focus the analysis on a more stable stage of stroke pathology, cases acquired during the acute and subacute phases were excluded. During these earlier stages, stroke lesions undergo dynamic biological and structural changes, including evolving edema, inflammatory responses, and tissue remodeling, resulting in substantial variability in lesion appearance and boundaries [Yuh et al., 1991, Cotinat et al., 2025, Hernandez Petzsche et al., 2022]. Restricting the dataset to chronic-stage cases, where these transient processes have largely resolved, provided a more consistent setting for evaluating the impact of different pre- training strategies on downstream stroke lesion segmenta- tion. After these exclusions, 649 chronic-stage scans re- mained for the experiments. T. Nasser et al.: Preprint submitted to ElsevierPage 3 of 16 Investigating Pretext Tasks for Brain MR Segmentation 2.2.4. Dataset for Cortical Brain Structure Segmentation The Mindboggle-101 dataset [Klein and Tourville, 2012] was used for the healthy brain cortical segmentation task. The dataset consists of 101 T1w MR images from healthy participants collected from multiple publicly available neu- roimaging datasets. The constituent datasets include the Nathan Kline Institute Rockland Sample (NKI-RS-22), Nathan Kline Institute Test–Retest dataset (NKI-TRT-20), Multi- Modal MRI Reproducibility Resource (MMRR-21), MMRR 3T/7T dataset (MMRR-3T7T-2), Human Language Network dataset (HLN-12), OASIS Test–Retest dataset (OASIS-TRT- 20), Colin27, and additional subjects included in the original Mindboggle-101 collection. The T1w MR images were processed using the FreeSurfer recon-all pipeline for cortical surface reconstruction and segmentation. Cortical regions were manually edited ac- cording to the Desikan–Killiany–Tourville (DKT) cortical labeling protocol, which was developed to provide anatom- ically consistent cortical parcellations based on reliable anatomical landmarks. The resulting manually labeled data provide reference cortical parcellations suitable for training and evaluating automated brain segmentation methods. 2.3. Preprocessing For the CNS dataset used in the paired T1w–FLAIR pretraining configuration, an additional within-subject mul- timodal registration step was performed prior to model train- ing. For each subject, the FLAIR image was rigidly regis- tered to the corresponding T1w image using FreeSurfer’s eA_robust_register, with the T1w image serving as the refer- ence. The estimated rigid transformation was then applied to resample the FLAIR image into T1w space, ensuring spatial correspondence between the two modalities [Reuter et al.]. For all datasets, MR scans were resampled to an isotropic resolution of 1 m³ and their intensities were normalized to the range [0,1]. Due to GPU memory limitations, all models were trained using randomly cropped image patches of size 128×128×128. 2.4. Model Architecture Three network architectures, U-Net, UNETR, and Swin- UNETR, were initially evaluated to identify the most suit- able backbone for the proposed framework [Ronneberger et al., 2015, Hatamizadeh et al., 2022, Tang et al., 2022]. U-Net relies on convolutional operations, whereas UNETR employs a Transformer-based encoder to capture global contextual information. SwinUNETR combines hierarchi- cal feature extraction with window-based self-attention, en- abling efficient modeling of both local anatomical details and long-range spatial dependencies. These characteristics make it particularly well suited for learning volumetric brain MR representations. Based on a preliminary comparison using the voxel-level brain age prediction task (Table 1), SwinUNETR achieved the best performance across both pre- training dataset configurations and was therefore selected as the backbone for all subsequent pretraining and downstream segmentation experiments. 2.5. Self-Supervised Pretraining Strategies To develop the proposed neuroimaging FM, SwinUNETR was pretrained using three self-supervised learning strate- gies: image inpainting (Section 2.5.1), voxel-level brain age prediction (Section 2.5.2), and a multitask framework (Section 2.5.3) that jointly optimizes both objectives. Each strategy was independently pretrained using the datasets illustrated in Figure 2 before being fine-tuned on the down- stream segmentation tasks. The pretraining input modalities were selected to match the requirements of the correspond- ing downstream task. Specifically, T1w MR was used for cortical brain structure segmentation and ischemic stroke lesion segmentation, whereas concatenated T1w and FLAIR images were used for MS lesion segmentation to maintain consistency between the pretraining and downstream input modalities. For each pretraining strategy, multiple optimiza- tion hyperparameter configurations were empirically evalu- ated during preliminary experiments. The best-performing configuration, selected based on the validation set, was adopted and is reported in the corresponding subsection. 2.5.1. Image Inpainting Image inpainting was selected as a non-domain specific self supervised pretext task because it is well established in the computer vision domain [Pathak et al., 2016]. The objective is to reconstruct missing regions of an input MR volume by exploiting the anatomical context of the sur- rounding voxels, thereby encouraging the model to learn meaningful structural representations that can be transferred to downstream tasks. Following previous work [Elharrouss et al., 2020, Kim et al., 2025, Tang et al., 2022], missing regions were sim- ulated using a coarse dropout strategy. Specifically, 12 cu- bic regions of size (32 × 32 × 32) were randomly re- moved/cropped from each input patch of size (128 × 128 × 128), masking approximately 19% of the input volume. Randomly positioned masks avoid the trivial solutions asso- ciated with fixed masking patterns while exposing the model to diverse anatomical contexts during training. The network was trained to reconstruct the original MR volume from the corrupted input. To improve reconstruc- tion quality beyond voxel-wise similarity, a perceptual loss based on AlexNet feature representations was incorporated, encouraging the reconstructed images to preserve high-level anatomical structures in addition to local image intensities [Zhang et al., 2018, Stoean et al., 2022]. The model was trained using the Adam optimizer with an initial learning rate of 1 × 10 −3 and a weight decay of 1×10 −4 . A StepLR scheduler was employed with a step size of 70 epochs and a decay factor of 훾 = 0.8. The network was optimized by minimizing the perceptual reconstruction loss. 2.5.2. Voxel-Level Brain Age Prediction Voxel-level brain age prediction is proposed as a domain- specific self-supervised pretext task in which the network learns to predict a 3D brain age map from the input MR volume. Unlike conventional brain age prediction, which T. Nasser et al.: Preprint submitted to ElsevierPage 4 of 16 Investigating Pretext Tasks for Brain MR Segmentation estimates a single chronological age for the entire brain, the proposed formulation assigns an age value to every voxel within the brain mask, encouraging the network to learn spatially localized age-related anatomical representations. To generate the training labels, each voxel within the brain mask was assigned the subject’s chronological age. During training, integer noise uniformly sampled from ([- 2,2]) years was added to the age labels to introduce voxel- wise label variability and prevent the network from converg- ing to the trivial solution of assigning the same age value to all voxels [Gianchandani et al., 2024]. The noise was applied only during training and was removed during validation and testing, where the original chronological age was used to provide an unbiased evaluation of the learned represen- tations. The relatively small noise range was selected to introduce sufficient variability while avoiding systematic bias in the age prediction task [Gianchandani et al., 2024]. The network was optimized using the Mean Absolute Error (MAE) loss, computed between the predicted and ground-truth voxel-wise age maps within the brain mask. By minimizing the voxel-wise MAE, the model learns age- sensitive representations that capture both local anatomical characteristics and global structural information relevant to brain aging. The model was trained using the Adam optimizer with an initial learning rate of 1 × 10 −3 , a weight decay of 1 × 10 −4 , and a StepLR scheduler with a step size of 150 epochs and a decay factor (훾) of 0.6. 2.5.3. Multitask Framework Building upon our previous work [Nasser et al., 2026], we further propose a multitask self-supervised pretraining framework that jointly optimizes voxel-level brain age pre- diction and image inpainting. The motivation is to learn richer and more transferable feature representations by com- bining a domain-specific pretext task, voxel-level brain age prediction, which encourages the network to learn anatomi- cally meaningful contextual and age-related representations, with a non-domain-specific pretext task, image inpainting, which learns contextual and structural representations by reconstructing missing image regions. By jointly optimizing both objectives, the shared SwinUNETR encoder is encour- aged to learn complementary representations that capture both domain-specific anatomical characteristics and general contextual image features. During training, the input MR volume is first masked using the same coarse dropout strategy employed for im- age inpainting. The same corrupted input is then provided to a shared SwinUNETR architecture, where both pretext tasks share the same encoder and decoder and differ only in their task-specific prediction heads. The reconstruction head predicts the original MR volume, while the brain age prediction head estimates the voxel-wise brain age map only for the visible (non-corrupted) brain regions. This design enables the shared network to simultaneously learn contex- tual structural representations through image reconstruction and contextual domain-specific anatomical representations, while preventing the brain age prediction task from relying on artificially masked regions. The multitask model is optimized using a weighted combination of the reconstruction and brain age prediction objectives: 퐿 total = 휆 inp 퐿 perc + 휆 age 퐿 MAE (1) where (퐿 perc ) denotes the perceptual reconstruction loss and (퐿 MAE ) is the voxel-wise MAE loss for brain age prediction. The weighting coefficients were empirically set to (휆 inp = 100) and (휆 age = 1). The multitask model was trained using the Adam opti- mizer with an initial learning rate of 1 × 10 −3 and a weight decay of 1 × 10 −4 . The learning rate was updated using a StepLR scheduler with a step size of 150 epochs and a decay factor of 훾 = 0.8. 2.6. Fine-tuning for the Downstream Segmentation Tasks The pretrained models are evaluated on three down- stream MR segmentation tasks. MS lesion segmentation and stroke lesion segmentation are considered non-age-related tasks, whereas cortical brain structure segmentation is se- lected as an age-related task. For each downstream segmentation task, multiple fine- tuning hyperparameter configurations were empirically eval- uated during preliminary experiments. The best-performing configuration, selected based on the validation set, was adopted and is reported in its corresponding subsection. To evaluate the effect of pretraining under different levels of downstream data availability, each task was further as- sessed using multiple training-set sizes. The selected subset sizes were constrained by the total number of available sub- jects in each dataset and were chosen to span progressively larger training regimes, beginning with the smallest feasible subset. This design enables evaluation of the relative benefit of pretraining compared with training from scratch in low- data settings and how this benefit changes as the amount of labeled downstream data increases. In particular, it allows us to examine whether the performance advantage of pretrained models decreases, disappears, or is potentially reversed as sufficient labeled data become available for effective training from scratch. 2.6.1. MS Lesion Segmentation Unlike the other downstream tasks, MS lesion segmen- tation requires both T1w and FLAIR MR, as MS lesions are often difficult to identify on T1w images alone and are better visualized on FLAIR. Consequently, pretraining for this task was performed using the paired T1w and FLAIR images from the CNS dataset. The downstream segmentation experiments were con- ducted using progressively increasing training set sizes of 11, 22, 30, 42, 50, and 58 subjects, while maintaining fixed validation and test sets of 20 and 15 subjects, respectively. For each training set size, four initialization strategies were T. Nasser et al.: Preprint submitted to ElsevierPage 5 of 16 Investigating Pretext Tasks for Brain MR Segmentation evaluated: training from scratch, image inpainting pretrain- ing, voxel-level brain age prediction pretraining, and the pro- posed multitask pretraining framework. All methods were fine-tuned using identical training, validation, and test splits to ensure a fair comparison. Different fine-tuning strategies were investigated for the MS lesion segmentation task. First, the entire pretrained network was fine-tuned using different learning rates (10 −3 , 10 −4 , and 10 −5 ). We then evaluated a strategy in which the segmentation decoder was reinitialized prior to end-to-end fine-tuning of the entire network. The decoder was reini- tialized to allow it to learn task-specific segmentation rep- resentations while retaining the transferable representations learned by the pretrained encoder. Additional experiments included freezing the pretrained encoder while optimizing only the reinitialized decoder and segmentation head, as well as a progressive fine-tuning strategy in which the decoder and segmentation head were first optimized before unfreez- ing the entire network for joint optimization. The model achieving the highest validation performance was selected, and the corresponding test results are reported. The best-performing fine-tuning strategy, in which the segmentation decoder was reinitialized prior to end-to-end fine-tuning of the entire network, was trained using the Adam optimizer with an initial learning rate of 1 × 10 −3 , a weight decay of 1 × 10 −4 , and a StepLR scheduler with a step size of 120 epochs and a decay factor (훾) of 0.6. Mod- els initialized from the self-supervised pretrained weights were fine-tuned for 300 epochs, whereas the model trained from scratch was trained for 400 epochs to allow sufficient convergence without the benefit of pretrained initialization. All other training settings were kept consistent across the evaluated strategies. 2.6.2. Ischemic Stroke Lesion Segmentation Stroke lesion segmentation was selected as an addi- tional downstream task unrelated to brain aging, providing a more independent assessment of the transferability of the learned representations. In the present study, this task was formulated using the T1w scans available in the ISLES 2026 dataset. The T1w-only input configuration enabled all three pretraining strategies to be trained using the larger multi-source normative pretraining dataset combined from several publicly available cohorts. This provided a substan- tially more diverse pretraining setting than the MS lesion segmentation experiments, for which the requirement for paired T1w and FLAIR images restricted pretraining to the CNS dataset. Consequently, the stroke lesion segmen- tation experiments provided a complementary assessment of whether representations learned from a larger and more heterogeneous pretraining cohort could generalize to a non- age-related downstream segmentation task. A fixed validation set of 50 scans and a fixed test set of 100 scans were used throughout all experiments to ensure direct comparability between the evaluated methods. From the remaining scans, nested training subsets containing 10, 20, 50, 100, 150, 200, 250, 300, 350, and 400 cases were constructed such that each smaller subset was fully con- tained within the subsequent larger subset. To ensure a fair comparison, identical training subsets were used across all training strategies at each dataset size. This experimental design simulated realistic scenarios with limited annotated data, where pretraining is expected to provide the greatest benefit, while ensuring that performance differences were attributable solely to the learning strategy rather than dif- ferences in the sampled training cohorts. The progressively increasing training subsets enabled a systematic evaluation of the effectiveness of pretraining under varying levels of la- beled data availability, with particular emphasis on the low- data regime and on how the benefits of pretraining evolve as additional annotated training data become available. The fine-tuning strategies evaluated for ischemic stroke lesion segmentation were selected based on the findings obtained from the MS lesion segmentation experiments. Specifically, all pretrained models, including voxel-level brain age prediction, image inpainting, and the proposed multitask framework, were fine-tuned by updating all net- work parameters. In addition, an alternative fine-tuning strat- egy was evaluated for the multitask framework, in which the decoder was randomly reinitialized while the encoder was initialized with the pretrained weights before fine-tuning the entire network. This strategy was included because it demon- strated best performance in the MS lesion segmentation experiments. Models trained from scratch were optimized using the same network architecture and training settings to ensure a fair comparison with their pretrained counterparts. All models were trained for 300 epochs using the Adam optimizer with an initial learning rate of 1 × 10 −3 , a weight decay of 1×10 −4 , and a StepLR learning-rate scheduler with a step size of 120 epochs and a decay factor (훾) of 0.6. An effective batch size of 8 was used for all experiments. The model achieving the lowest validation loss was selected for evaluation on the independent test set. Each experimental configuration, defined by the combination of training-set size and training strategy, was repeated five times using different random seeds to account for stochastic variation during training. All reported quantitative results are pre- sented as the mean± standard deviation across the five independent runs. 2.6.3. Cortical Brain Structure Segmentation Cortical brain structure segmentation was selected as an age-related downstream task to evaluate whether rep- resentations learned through voxel-level brain age predic- tion transfer effectively to the segmentation of anatomical structures that undergo systematic changes across the adult lifespan. In contrast to the MS and ischemic stroke lesion segmentation tasks, which were selected as downstream applications not directly related to brain aging, cortical seg- mentation provided a complementary setting for investi- gating the transferability of the learned representations to healthy brain anatomy. The task was formulated using the T1w scans and corresponding DKT cortical parcellations from the Mindboggle-101 dataset. As this task requires only T. Nasser et al.: Preprint submitted to ElsevierPage 6 of 16 Investigating Pretext Tasks for Brain MR Segmentation T1w images, all three pretraining strategies were initialized using the larger multi-source normative T1w pretraining dataset, enabling direct comparison of voxel-level brain age prediction, image inpainting, and the proposed multitask framework under the same pretraining-data configuration used for ischemic stroke lesion segmentation. A fixed validation set of 15 scans and a fixed test set of 20 scans were used throughout all experiments to ensure direct comparability between the evaluated methods. From the remaining scans, nested training subsets containing 10, 25, 45, and 65 cases were constructed such that each smaller subset was fully contained within the subsequent larger sub- set. Identical training subsets were used across all training strategies at each dataset size to ensure that differences in segmentation performance were attributable to the learning strategy rather than differences in the sampled training co- horts. The progressively increasing training subsets enabled a systematic evaluation of the effectiveness of pretraining under different levels of labeled data availability and allowed us to investigate whether the relative benefit of the different pretraining strategies changed as additional annotated data became available. The fine-tuning strategies evaluated for cortical brain structure segmentation followed those identified as the most effective in the preceding downstream experiments. Specifi- cally, the voxel-level brain age prediction and image inpaint- ing pretrained models were fine-tuned by updating all net- work parameters. The proposed multitask pretrained model was evaluated using two fine-tuning strategies: fine-tuning all pretrained network parameters and randomly reinitial- izing the decoder while retaining the pretrained encoder weights before fine-tuning the entire network. A model trained from scratch using the same network architecture was included as the baseline to provide a direct comparison with the pretrained models. All models were trained for 300 epochs using the Adam optimizer with an initial learning rate of 1 × 10 −3 , a weight decay of 1×10 −4 , and a StepLR learning-rate scheduler with a step size of 150 epochs. For the model pretrained using voxel-level brain age prediction, an additional learning-rate scheduling configuration with a reduced step size of 100 epochs was evaluated to investigate whether more frequent learning-rate decay could improve downstream fine-tuning performance. However, the original configuration with a step size of 150 epochs consistently provided better performance and was therefore retained for the final comparison and reported results. An effective batch size of 8 was used for all experiments. The model achieving the lowest validation loss was selected for evaluation on the test set. Each experimental configuration, defined by the combination of training-set size and training strategy, was repeated five times using different random seeds to account for stochastic variation during training. All reported quantitative results are pre- sented as the mean ± standard deviation across the five independent runs. Table 1 Backbone architecture comparison for the voxel-level brain age prediction pretext task. Lower MAE indicates better performance. Architecture Multi-source T1w CNS (T1w+FLAIR) MAE (years)MAE (years) U-Net6.4 ± 4.36.9 ± 4.7 UNETR7.1 ± 4.48.0 ± 5.0 SwinUNETR6.1 ± 4.15.3 ± 3.1 3. Results 3.1. Self-Supervised Pretraining Strategies To identify the most suitable backbone architecture for the proposed voxel-level brain age prediction framework, an ablation study was conducted using three state-of-the- art architectures: U-Net, UNETR, and SwinUNETR. The comparison was performed under two pretraining settings: (1) using T1w MR from the multi-source public dataset, and (2) using concatenated T1w and FLAIR MR from CNS dataset. As shown in Table 1, SwinUNETR consistently achieved the lowest MAE in both settings and was therefore selected as the backbone architecture for all subsequent pretraining and downstream segmentation experiments. Table 2 summarizes the performance of the three self- supervised pretraining strategies under the two pretrain- ing settings. Compared with the corresponding single-task frameworks, the proposed multitask model achieved im- proved reconstruction performance and voxel-level brain age prediction accuracy. This suggests that the two pretext tasks provide complementary supervisory signals, motivating the evaluation of whether the jointly learned representations improve performance on downstream segmentation tasks. 3.2. Fine-tuning on Downstream Segmentation Tasks 3.2.1. Multiple Sclerosis Lesion Segmentation To determine the most effective transfer learning strat- egy, multiple fine-tuning configurations were initially evalu- ated for each pretrained model, together with a model trained from scratch for comparison. Based on these experiments, fine-tuning the entire network while reinitializing the seg- mentation decoder consistently achieved the best perfor- mance for the pretrained models and was therefore selected for the final comparison. Table 3 summarizes the quantita- tive Dice scores obtained using the best-performing fine- tuning strategy, while Figure 3 illustrates the correspond- ing performance variability across the different training set sizes. Among the evaluated pretraining strategies, the proposed multitask framework generally achieved the highest Dice scores, outperforming both image inpainting and voxel- level brain age prediction across most training set sizes. Compared with training from scratch, multitask pretraining provided the largest improvements when only a very limited number of annotated training samples were available (11 subjects) and continued to outperform training from scratch T. Nasser et al.: Preprint submitted to ElsevierPage 7 of 16 Investigating Pretext Tasks for Brain MR Segmentation Table 2 Performance of the three self-supervised pretraining strategies under the two pretraining settings. Lower perceptual loss and voxel MAE indicate better performance, whereas higher PSNR indicates better reconstruction quality. Pretraining StrategyPerceptual Loss PSNR (dB) Voxel MAE (years) Image Inpainting Multi-source T1w 0.024034.90– CNS (T1w+FLAIR)0.003343.52– Voxel-Level Brain Age Multi-source T1w–6.1 ± 4.1 CNS (T1w+FLAIR) –5.3 ± 3.1 Multitask Multi-source T1w0.002046.245.46 ± 3.7 CNS (T1w+FLAIR)0.002645.785.25 ± 3.1 Table 3 Comparison of different fine-tuning strategies for MS lesion segmentation. For all pretrained models, the segmentation decoder was reinitialized prior to fine-tuning. Each experiment was independently repeated five times, and the mean Dice score ± standard deviation is reported. The best result for each training set is highlighted in bold italic, while bold indicates results within a Dice difference of 0.005 from the best-performing method. TrainingScratchVoxel-Level Brain AgeImage InpaintingMultitask SizeEnd-to-EndAll LayersDecoder OnlyAll LayersDecoder OnlyAll LayersDecoder Only 110.662 ± 0.0160.631 ± 0.013 0.593 ± 0.0060.654 ± 0.022 0.623 ± 0.0080.671 ± 0.016 0.654 ± 0.014 22 0.699 ± 0.0150.692 ± 0.012 0.685 ± 0.0170.696 ± 0.019 0.661 ± 0.0090.695 ± 0.025 0.699 ± 0.010 300.751 ± 0.0130.715 ± 0.015 0.722 ± 0.0100.727 ± 0.012 0.700 ± 0.0090.737 ± 0.006 0.706 ± 0.002 420.757 ± 0.0170.733 ± 0.014 0.724 ± 0.0140.743 ± 0.010 0.725 ± 0.0130.752 ± 0.012 0.727 ± 0.005 500.757 ± 0.0170.749 ± 0.011 0.736 ± 0.0140.747 ± 0.012 0.724 ± 0.0100.764 ± 0.010 0.741 ± 0.009 580.764 ± 0.0190.763 ± 0.008 0.750 ± 0.0100.765 ± 0.008 0.732 ± 0.0140.775 ± 0.006 0.751 ± 0.002 at the larger training set sizes (50 and 58 subjects). For the 22- and 42-subject training sets, the two approaches achieved comparable performance, whereas training from scratch produced the highest Dice score for the 30-subject training set. Although the available dataset does not allow us to determine whether this trend would persist with sub- stantially larger training sets, these findings suggest that the proposed multitask pretraining is particularly beneficial in low-data settings while remaining competitive as more annotated training data become available. Qualitative results for a representative test case in the 11-subject fine-tuning setting are shown in Fig. 4. This motivated us to further evaluate the proposed framework on ischemic stroke lesion segmentation, which requires only T1w MR. 3.2.2. Ischemic Stroke Lesion Segmentation Table 4 summarizes the performance of the evaluated training strategies for ischemic stroke lesion segmentation across different training-set sizes, reporting the mean Dice score (± standard deviation) over five independent runs for each method. As shown in Table 4, the pretrained models generally outperform the model trained from scratch in data-scarce scenarios. The proposed multitask pretraining framework consistently achieves the highest segmentation performance for training-set sizes up to 150 labeled cases. Although the relative ranking of the two multitask fine-tuning strategies varies slightly across individual training-set sizes, both con- sistently outperform the individual pretraining strategies and the randomly initialized baseline. Specifically, fine-tuning all pretrained layers provides the highest performance for the 10- and 20-case settings, whereas decoder reinitialization followed by end-to-end fine-tuning yields the best perfor- mance for the 50- and 100-case settings. At 150 labeled cases, the two multitask strategies achieve comparable Dice scores, indicating that no single fine-tuning strategy is uni- formly optimal across the low-data regime. The results in Table 4 also demonstrate that the benefit of pretraining is most pronounced when labeled data are scarce. The proposed multitask framework provides the largest per- formance gains over training from scratch, reaching im- provements of approximately 6–7 Dice percentage points for the smallest training subsets. As the number of labeled training cases increases, the performance gap between the pretrained and randomly initialized models progressively decreases, becoming minimal beyond approximately 200 training cases. This trend indicates that pretraining is partic- ularly beneficial in the low-data regime, whereas its relative advantage diminishes as more labeled data become available for supervised training. Another notable observation is the difference in perfor- mance stability among the evaluated pretraining strategies. Beyond approximately 200 labeled training cases, the image- inpainting approach exhibits larger fluctuations in mean Dice across consecutive training-set sizes, suggesting greater sen- sitivity to the amount or composition of the labeled training data. In contrast, both multitask variants show more sta- ble performance as the training-set size increases. Among them, fine-tuning all pretrained layers exhibits the most consistent behavior across the evaluated training-set sizes, whereas decoder reinitialization shows somewhat greater variability while remaining more stable than the individual pretraining approaches. To further examine performance at the individual-case level, Fig. 5 shows the per-image Dice scores for all 100 test cases from the fourth independent T. Nasser et al.: Preprint submitted to ElsevierPage 8 of 16 Investigating Pretext Tasks for Brain MR Segmentation Figure 3: Comparison of the best-performing fine-tuning strategy for the MS lesion segmentation task, in which the entire network was fine-tuned after reinitializing the segmentation decoder. Table 4 Mean Dice score ± standard deviation over five independent runs for ischemic stroke lesion segmentation across different training- set sizes and training strategies. Training SizeScratchBrain AgeInpaintingMultitask–Decoder Reinitialized Multitask–FT All 100.216 ± 0.033 0.203 ± 0.010 0.206 ± 0.0280.276 ± 0.0270.286 ± 0.011 200.229 ± 0.010 0.249 ± 0.010 0.231 ± 0.0240.283 ± 0.0300.300 ± 0.030 500.332 ± 0.025 0.351 ± 0.016 0.329 ± 0.0230.399 ± 0.0160.391 ± 0.009 1000.459 ± 0.017 0.459 ± 0.009 0.464 ± 0.0160.480 ± 0.0140.470 ± 0.012 1500.488 ± 0.009 0.485 ± 0.009 0.488 ± 0.0150.481 ± 0.0240.492 ± 0.011 2000.506 ± 0.010 0.506 ± 0.008 0.507 ± 0.0220.525 ± 0.0080.502 ± 0.031 2500.526 ± 0.004 0.517 ± 0.019 0.502 ± 0.0170.522 ± 0.0080.536 ± 0.001 3000.530 ± 0.013 0.533 ± 0.013 0.525 ± 0.0130.520 ± 0.0090.534 ± 0.006 3500.539 ± 0.010 0.546 ± 0.005 0.524 ± 0.0170.536 ± 0.0110.533 ± 0.005 4000.537 ± 0.011 0.539 ± 0.008 0.533 ± 0.0140.542 ± 0.0130.540 ± 0.009 run using 50 labeled training cases. The figure highlights the substantial inter-case variability in segmentation perfor- mance while enabling direct comparison of the evaluated strategies on the same test images. A qualitative comparison for a representative test case from the same 50-case fine- tuning setting is provided in Fig. 6. The same test image and anatomical slices are shown across all evaluated strategies, with the ground-truth lesion in green and the corresponding predictions in violet, enabling direct visual comparison of lesion localization and extent. Overall, these findings indicate that combining voxel- level brain age prediction and image inpainting within a unified multitask pretraining framework yields representa- tions that transfer more effectively to downstream ischemic stroke lesion segmentation than either pretext task individ- ually. The advantage of multitask pretraining is particu- larly evident in the low-data regime, while the differences among training strategies become progressively smaller as the amount of labeled training data increases. T. Nasser et al.: Preprint submitted to ElsevierPage 9 of 16 Investigating Pretext Tasks for Brain MR Segmentation Figure 4: MS lesion segmentation results for the same representative test image across the evaluated pretraining strategies with 11 images used for fine-tuning. Although both T1w and FLAIR images were used as model inputs, the segmentation results are overlaid on the FLAIR image for clearer visualization of the lesions. For all pretrained models, the decoder was reinitialized before fine-tuning; all layers were subsequently fine-tuned except for the Multitask decoder-only setting, in which only the reinitialized decoder was trained. Table 5 Mean Dice score ± standard deviation over five independent runs for cortical brain structure segmentation across different training-set sizes and training strategies. Training SizeScratchBrain AgeInpaintingMultitask–FT All Multitask–Decoder Reinitialized 100.186 ± 0.023 0.063 ± 0.010 0.221 ± 0.024 0.248 ± 0.0330.251 ± 0.049 250.493 ± 0.063 0.313 ± 0.047 0.545 ± 0.080 0.633 ± 0.0440.611 ± 0.033 450.726 ± 0.035 0.723 ± 0.019 0.762 ± 0.0050.732 ± 0.0410.674 ± 0.060 650.741 ± 0.016 0.737 ± 0.036 0.772 ± 0.0150.762 ± 0.0160.742 ± 0.033 3.2.3. Cortical Brain Structure segmentation Table 5 summarizes the cortical brain structure segmen- tation performance across the four training-set sizes. The proposed multitask pretraining framework outperformed training from scratch across all evaluated training-set sizes, with the largest improvements observed in the low-data regime. At 10 training cases, Multitask–Decoder Reini- tialized achieved the highest mean Dice score of 0.251 ± 0.049, compared with 0.186 ± 0.023 for training from scratch. Multitask–FT All achieved a similar Dice score of 0.248 ± 0.033. At 25 training cases, Multitask–FT All achieved the highest performance (0.633 ± 0.044), fol- lowed by Multitask–Decoder Reinitialized (0.611 ± 0.033), compared with 0.493 ± 0.063 for training from scratch. A representative qualitative example from the same 25-case fine-tuning setting is shown in Fig. 7. The same test image slices are displayed across the evaluated strategies, enabling direct visual comparison of the predicted cortical regions. As the number of labeled training cases increased, the relative performance of the evaluated pretraining strategies changed. At 45 and 65 training cases, image inpainting achieved the highest mean Dice scores, reaching 0.762 ± 0.005 and 0.772 ± 0.015, respectively. At these training- set sizes, Multitask–FT All remained competitive, achiev- ing Dice scores of 0.732 ± 0.041 and 0.762 ± 0.016, re- spectively, and continued to outperform the corresponding models trained from scratch. In contrast, voxel-level brain age prediction alone did not provide a consistent advantage over random initialization for cortical brain structure segmentation. The brain age pre- trained model achieved lower mean Dice scores than training from scratch at all four training-set sizes, with the difference being particularly pronounced in the low-data settings. At T. Nasser et al.: Preprint submitted to ElsevierPage 10 of 16 Investigating Pretext Tasks for Brain MR Segmentation 0.0 0.2 0.4 0.6 0.8 1.0 Dice(c1) Scratch Mean Dice = 0.2888 0.0 0.2 0.4 0.6 0.8 1.0 Dice(c1) Multitask FT All Mean Dice = 0.3925 0.0 0.2 0.4 0.6 0.8 1.0 Dice(c1) Inpainting Mean Dice = 0.3065 0.0 0.2 0.4 0.6 0.8 1.0 Dice(c1) Brain Age Mean Dice = 0.3548 020406080100 Test Sample Number 0.0 0.2 0.4 0.6 0.8 1.0 Dice(c1) Multitask Decoder Reinitialized Mean Dice = 0.3874 Per-Image Dice Scores for the Five Training Techniques Data Center sub-r001 sub-r002 sub-r003 sub-r004 sub-r005 sub-r009 sub-r017 sub-r018 sub-r024 sub-r027 sub-r028 sub-r029 sub-r030 sub-r032 sub-r034 sub-r035 sub-r038 sub-r040 sub-r042 sub-r046 sub-r047 sub-r048 sub-r052 sub-r056 sub-r057 sub-r061 sub-r062 sub-r067 Figure 5: Per-image Dice scores for the five training strategies on the ischemic stroke lesion segmentation task. Results are shown for the fourth independent run using 50 labeled training cases. Each point represents one of the 100 test images, displayed in the same order across all subplots to enable direct comparison between methods. Colors indicate the acquisition center of each test image, and the dashed horizontal line denotes the mean Dice score. The evaluated strategies are training from scratch, image inpainting pretraining, voxel-level brain age prediction pretraining, multitask pretraining with fine-tuning of all pretrained layers (Multitask–FT All), and multitask pretraining with decoder reinitialization followed by end-to-end fine-tuning (Multitask–Decoder Reinitialized). 10 training cases, brain age pretraining achieved a mean Dice score of 0.063 ± 0.010, compared with 0.186 ± 0.023 for training from scratch. Similarly, at 25 cases, the corre- sponding Dice scores were 0.313±0.047 and 0.493±0.063, respectively. The difference became substantially smaller as the training-set size increased, with brain age pretraining and training from scratch achieving comparable performance at 45 and 65 cases. Overall, multitask pretraining achieved the highest performance in the low-data settings, whereas image inpainting performed best with the larger training subsets, while voxel-level brain age pretraining alone did not outperform training from scratch. 4. Discussion and Future Perspectives In this study, we investigated voxel-level brain age pre- diction as a domain-specific self-supervised pretext task and compared it with image inpainting, a widely used non- domain-specific self-supervised objective. Building upon these complementary pretext tasks, we proposed a multi- task self-supervised learning framework that jointly opti- mizes both objectives to learn richer and more transferable neuroimaging representations. Rather than optimizing the learned representations for a single downstream application, our objective was to develop a pretraining strategy capable of generalizing across diverse neurological tasks, regardless of their direct relationship to brain aging. T. Nasser et al.: Preprint submitted to ElsevierPage 11 of 16 Investigating Pretext Tasks for Brain MR Segmentation Figure 6: Stroke lesion segmentation results for the same representative test image across the evaluated pretraining strategies with 50 images used for fine-tuning in a single run. The segmentation results are overlaid on the T1w image, with the ground-truth lesion shown in green and the predicted lesions shown in violet. To evaluate this hypothesis, we considered three down- stream segmentation tasks with distinct characteristics: cor- tical brain structure segmentation, representing an anatomi- cal task involving structures known to be affected by aging, and MS and ischemic stroke lesion segmentation, repre- senting two clinically relevant pathological tasks not di- rectly related to the normal brain aging process. This ex- perimental design enabled us to investigate whether rep- resentations learned through voxel-level brain age predic- tion preferentially benefit anatomically age-sensitive appli- cations, whether the more generic image inpainting ob- jective provides broader transferability, or whether jointly optimizing both objectives produces complementary repre- sentations that are particularly advantageous when labeled downstream data are scarce. Overall, the proposed multitask framework achieved the strongest or among the strongest performance across the evaluated downstream tasks, with its advantage being par- ticularly evident in several low-data settings. These findings suggest that combining domain-specific and non-domain- specific self-supervision can provide more transferable rep- resentations than relying exclusively on either objective. Unlike the other downstream tasks, MS lesion segmen- tation requires both T1w and FLAIR MR images because MS lesions are often difficult to identify on T1w images alone and are more conspicuous on FLAIR. Consequently, pretraining for this task was restricted to the CNS dataset, which provides paired T1w and FLAIR images together with chronological age information. Neither image inpainting nor voxel-level brain age prediction alone consistently outper- formed training from scratch, whereas their combination within the proposed multitask framework generally yielded the strongest performance. One factor specific to the MS experiments is the limited diversity of the available pretraining data. Because paired T1w and FLAIR images were required, pretraining was restricted to the single-source CNS cohort. Although this en- sured consistency between the pretraining and downstream input modalities, the limited variation in imaging protocols, scanners, and subject populations may have constrained the transferability of the learned representations. A larger multi- source normative dataset containing paired T1w and FLAIR images together with chronological age information would enable a more comprehensive evaluation of this hypothesis and help distinguish limitations associated with the pretext tasks themselves from those arising from the scale and diversity of the available pretraining data. Although multitask pretraining generally outperformed training from scratch in the MS experiments, this trend was not observed for the training subset containing 30 subjects, where the randomly initialized model achieved the highest Dice score. Given the relatively small size of the MS lesion segmentation dataset, it remains unclear whether this result reflects statistical variability associated with the sampled training data or a genuine characteristic of the learning process. This observation motivated the evaluation of the proposed framework on an additional non-age-related patho- logical task, ischemic stroke lesion segmentation. T. Nasser et al.: Preprint submitted to ElsevierPage 12 of 16 Investigating Pretext Tasks for Brain MR Segmentation Figure 7: Cortical parcellation results for the same representative test image across the evaluated pretraining strategies with 25 images used for fine-tuning. In contrast to the MS experiments, ischemic stroke le- sion segmentation enabled all pretraining strategies to use the larger multi-source T1w normative dataset. This sub- stantially increased the diversity of the pretraining data by incorporating scans from multiple publicly available cohorts with differences in imaging protocols, scanners, and subject populations. The stroke experiments therefore provided an opportunity to evaluate the proposed framework under a substantially broader pretraining setting. The stroke lesion segmentation results demonstrated a clearer advantage of self-supervised pretraining than that observed for MS lesion segmentation. The benefit was most pronounced when labeled downstream data were limited, al- though its magnitude varied among the evaluated strategies. The proposed multitask framework achieved the highest segmentation performance for training-set sizes up to 150 labeled cases. Fine-tuning all pretrained layers yielded the highest Dice scores with 10 and 20 training cases, whereas decoder reinitialization followed by end-to-end fine-tuning performed best with 50 and 100 cases. At 150 cases, the two multitask fine-tuning strategies achieved comparable performance. The largest gains over training from scratch reached approximately 6–7 absolute Dice percentage points in the smallest training subsets. As the number of labeled cases increased, however, the performance difference be- tween pretrained and randomly initialized models progres- sively diminished. This pattern indicates that the value of the learned initialization is greatest when downstream super- vision is limited and becomes less pronounced as sufficient task-specific annotations become available. Differences in performance stability were also observed among the individual pretraining strategies. Image inpaint- ing exhibited greater variation as the number of labeled training cases increased, suggesting greater sensitivity to the composition of the downstream training set. Voxel-level brain age prediction generally showed more stable behavior, although it did not consistently outperform training from scratch. The multitask framework, in comparison, combined strong performance with relatively stable behavior across the evaluated training-set sizes. The per-image analysis further demonstrated that the improvements obtained with multitask pretraining were ob- served across images acquired from multiple centers rather than being driven by a small subset of test cases. Higher Dice scores were observed across a broad range of subjects and acquisition centers, providing additional evidence that the improvements were not restricted to a particular center or imaging condition. The cortical brain structure segmentation experiments revealed a different pattern from the lesion segmentation tasks. Multitask pretraining achieved the strongest perfor- mance in the low-data settings of 10 and 25 training cases, whereas image inpainting achieved the highest performance with 45 and 65 labeled cases. Unexpectedly, voxel-level brain age prediction alone did not outperform training from scratch at any of the evaluated training-set sizes, despite cortical anatomy being affected by the aging process. One possible explanation is that, although cortical mor- phology undergoes age-related changes, cortical parcellation is not itself a directly age-prediction-related task. Cortical T. Nasser et al.: Preprint submitted to ElsevierPage 13 of 16 Investigating Pretext Tasks for Brain MR Segmentation parcellation primarily requires the identification of anatomi- cal boundaries and spatial relationships among neighboring cortical regions, whereas voxel-level brain age prediction encourages the network to identify features associated with normative aging trajectories. Features that are informative for estimating chronological age may therefore not neces- sarily be those that are most discriminative for delineating cortical regions. The stronger performance of image inpainting with the larger cortical training subsets may also reflect the nature of the downstream task. Image inpainting requires recon- struction of missing image regions from their surrounding anatomical context, encouraging the model to learn local structural information, spatial relationships, and anatomical organization. Such representations may be particularly rel- evant to cortical parcellation, where accurate segmentation depends heavily on distinguishing neighboring anatomical structures. As additional labeled downstream data became available, these generic structural representations may have provided a more suitable initialization for supervised cortical parcellation than representations optimized specifically for brain age prediction. Nevertheless, the cortical experiments also provide an important observation regarding the proposed multitask for- mulation. Despite the limited performance of voxel-level brain age prediction when used independently, combining it with image inpainting substantially improved performance in the smallest training subsets. At 10 and 25 training cases, both multitask variants outperformed training from scratch and the individual pretraining strategies. The utility of the brain age objective should therefore not be assessed ex- clusively from its performance as an isolated pretext task. Within the joint optimization framework, age-related su- pervision may complement the structural and contextual information learned through image reconstruction, resulting in a more informative representation when downstream su- pervision is particularly limited. Taken together, the three downstream experiments in- dicate that the effectiveness of self-supervised pretraining depends on the interaction among the pretext objective, downstream task, amount of labeled downstream data, and diversity of the pretraining cohort. No individual pretext task was uniformly optimal across all experimental conditions. In contrast, the proposed multitask framework demonstrated a particular advantage in several low-data settings, supporting the central motivation of combining domain-specific age- related supervision with a generic reconstruction objective to capture complementary aspects of brain MR images. Several limitations should be considered when interpret- ing these findings. First, identifying an appropriate publicly available dataset with high-quality manual annotations for a downstream task directly related to brain aging was chal- lenging. Although cortical anatomy is affected by aging, cor- tical parcellation primarily evaluates anatomical boundary delineation and therefore represents only an indirect assess- ment of the transferability of age-related representations. Future work should evaluate the proposed framework on downstream tasks more directly associated with aging and age-related structural changes. Second, the present evaluation was restricted to segmen- tation tasks. Although the three selected applications pro- vide substantially different segmentation settings, including healthy anatomy and two pathological lesion segmentation problems, they do not fully establish the generalizability of the learned representations across other types of neuroimag- ing applications. Future work should therefore extend the evaluation to classification and prediction tasks, including neurological disease classification, cognitive or clinical out- come prediction, and other applications in which age-related representations may provide complementary information. Finally, the MS experiments were constrained by the availability of multimodal normative pretraining data. Be- cause MS lesion segmentation required paired T1w and FLAIR images, pretraining was restricted to a single-source cohort, whereas the T1w-only stroke and cortical exper- iments benefited from a larger multi-source pretraining dataset. Future work should investigate large-scale, multi- center normative datasets containing paired T1w and FLAIR images and chronological age information to determine whether increasing pretraining diversity yields similar im- provements for multimodal downstream tasks. 5. Conclusion This study investigated voxel-level brain age prediction as a domain-specific self-supervised pretext task and intro- duced a multitask framework that combines it with image inpainting to learn more transferable brain MR imaging rep- resentations. Across three downstream segmentation tasks, the proposed multitask approach demonstrated its greatest benefit in low-data settings, highlighting the value of com- bining complementary domain-specific and generic super- visory signals when annotated data are limited. The find- ings also indicate that the effectiveness of self-supervised pretraining is influenced by the scale and diversity of the pretraining data, with stronger transfer observed when a larger multi-source pretraining cohort was available. Over- all, these results support multitask self-supervised learn- ing as a promising approach for improving representation transfer in brain MR imaging, particularly in data scarce scenarios. 6. Acknowledgments This work was supported by research funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Canada Research Chairs Program. We also acknowledge the Digital Research Alliance of Canada for providing computational infrastructure support. References Ischemic stroke lesion segmentation challenge 2026 - grand challenge. URL https://isles-26.grand-challenge.org/dataset/. John Absher, Sarah Goncher, Roger Newman-Norlund, Nicholas Perkins, Grigori Yourganov, Jan Vargas, Sanjeev Sivakumar, Naveen Parti, T. Nasser et al.: Preprint submitted to ElsevierPage 14 of 16 Investigating Pretext Tasks for Brain MR Segmentation Shannon Sternberg, Alex Teghipco, Makayla Gibson, Sarah Wilson, Leonardo Bonilha, and Chris Rorden. The stroke outcome optimization project: Acute ischemic strokes from a comprehensive stroke center. 11: 839, 2024. ISSN 2052-4463. doi: 10.1038/s41597-024-03667-5. URL https://pmc.ncbi.nlm.nih.gov/articles/PMC11297183/. Maëva Cotinat, Noëlle Messaoudi, Emmanuelle Robinet, Laurent Suissa, Emilie Doche, Maxime Guye, Bertrand Audoin, Laurent Bensoussan, Jean-Philippe Ranjeva, and Wafaa Zaaraoui. Dynamics of ionic and cytotoxic edema during acute and subacute stages of patients with ischemic stroke: Complementarity of 23na MRI and diffusion MRI. 38 (5):e70028, 2025. ISSN 1099-1492. doi: 10.1002/nbm.70028. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/nbm.70028. A. Di Martino, C.-G. Yan, Q. Li, E. Denio, F. X. Castellanos, K. Alaerts, J. S. Anderson, M. Assaf, S. Y. Bookheimer, M. Dapretto, B. Deen, S. Delmonte, I. Dinstein, B. Ertl-Wagner, D. A. Fair, L. Gallagher, D. P. Kennedy, C. L. Keown, C. Keysers, J. E. Lainhart, C. Lord, B. Luna, V. Menon, N. J. Minshew, C. S. Monk, S. Mueller, R.-A. Müller, M. B. Nebel, J. T. Nigg, K. O’Hearn, K. A. Pelphrey, S. J. Peltier, J. D. Rudie, S. Sunaert, M. Thioux, J. M. Tyszka, L. Q. Uddin, J. S. Verhoeven, N. Wenderoth, J. L. Wiggins, S. H. Mostofsky, and M. P. Milham. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism. Molecular Psychiatry, 19(6): 659–667, June 2014. ISSN 1476-5578. doi: 10.1038/mp.2013.78. URL https://w.nature.com/articles/mp201378. A. Di Martino, David O’Connor, Bosi Chen, Kaat Alaerts, Jeffrey S. An- derson, Michal Assaf, Joshua H. Balsters, Leslie Baxter, Anita Beggiato, Sylvie Bernaerts, Laura M. E. Blanken, Susan Y. Bookheimer, B. Blair Braden, Lisa Byrge, F. Xavier Castellanos, Mirella Dapretto, Richard Delorme, Damien A. Fair, Inna Fishman, Jacqueline Fitzgerald, Louise Gallagher, R. Joanne Jao Keehn, Daniel P. Kennedy, Janet E. Lainhart, Beatriz Luna, Stewart H. Mostofsky, Ralph-Axel Müller, Mary Beth Nebel, Joel T. Nigg, Kirsten O’Hearn, Marjorie Solomon, Roberto Toro, Chandan J. Vaidya, Nicole Wenderoth, Tonya White, R. Cameron Craddock, Catherine Lord, Bennett Leventhal, and Michael P. Milham. Enhancing studies of the connectome in autism using the autism brain imaging data exchange I. Scientific Data, 4(1):170010, March 2017. ISSN 2052-4463. doi: 10.1038/sdata.2017.10. URL https://w. nature.com/articles/sdata201710. Omar Elharrouss, Noor Almaadeed, Somaya Al-Maadeed, and Younes Akbari. Image Inpainting: A Review. Neural Processing Letters, 51(2):2007–2028, April 2020. ISSN 1573-773X. doi: 10.1007/s11063-019-10163-0.URL https://doi.org/10.1007/ s11063-019-10163-0. Neha Gianchandani, Johanna Ospel, Ethan MacDonald, and Roberto Souza. A Multitask Deep Learning Model for Voxel-Level Brain Age Esti- mation. In Xiaohuan Cao, Xuanang Xu, Islem Rekik, Zhiming Cui, and Xi Ouyang, editors, Machine Learning in Medical Imaging, pages 283–292, Cham, 2024. Springer Nature Switzerland. ISBN 978-3-031- 45676-3. doi: 10.1007/978-3-031-45676-3_29. Francesco Guarnera, Alessia Rondinella, Elena Crispino, Giulia Russo, Clara Di Lorenzo, Davide Maimone, Francesco Pappalardo, and Sebas- tiano Battiato. MSLesSeg: baseline and benchmarking of a new Multiple Sclerosis Lesion Segmentation dataset. Scientific Data, 12(1):920, May 2025. ISSN 2052-4463. doi: 10.1038/s41597-025-05250-y. URL https://doi.org/10.1038/s41597-025-05250-y. Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R. Roth, and Daguang Xu. UNETR: Transformers for 3D Medical Image Segmentation. pages 574–584, 2022.URL https://openaccess.thecvf.com/content/ WACV2022/html/Hatamizadeh_UNETR_Transformers_for_3D_Medical_Image_ Segmentation_WACV_2022_paper.html. Sheng He, P Ellen Grant, and Yangming Ou. Global-local transformer for brain age estimation. IEEE transactions on medical imaging, 41(1):213– 224, 2021. Yuting He, Fuxiang Huang, Xinrui Jiang, Yuxiang Nie, Minghao Wang, Jiguang Wang, and Hao Chen. Foundation Model for Advancing Healthcare: Challenges, Opportunities and Future Directions. IEEE Reviews in Biomedical Engineering, 18:172–191, 2025. ISSN 1941- 1189. doi: 10.1109/RBME.2024.3496744. URL https://ieeexplore. ieee.org/document/10750441/. Moritz R. Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Valenzuela, Mauricio Reyes, Maria Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, David Robben, Alexandre Hutton, Tassilo Friedrich, Teresa Zarth, Johannes Bürkle, The Anh Baran, Björn Menze, Gabriel Broocks, Lukas Meyer, Claus Zimmer, Tobias Boeckh- Behrens, Maria Berndt, Benno Ikenberg, Benedikt Wiestler, and Jan S. Kirschke. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. 9(1):762, 2022. ISSN 2052- 4463. doi: 10.1038/s41597-022-01875-5. URL https://w.nature. com/articles/s41597-022-01875-5. IXI Dataset.IXI Dataset – Brain Development http://brain- development.org/ixi-dataset/. URL http://brain-development.org/ ixi-dataset/. Clifford R. Jack, Matt A. Bernstein, Nick C. Fox, Paul Thompson, Gene Alexander, Danielle Harvey, Bret Borowski, Paula J. Britson, Jennifer L. Whitwell, Chadwick Ward, Anders M. Dale, Joel P. Felmlee, Jef- frey L. Gunter, Derek L.G. Hill, Ron Killiany, Norbert Schuff, Sab- rina Fox-Bosetti, Chen Lin, Colin Studholme, Charles S. DeCarli, Gunnar Krueger, Heidi A. Ward, Gregory J. Metzger, Katherine T. Scott, Richard Mallozzi, Daniel Blezek, Joshua Levy, Josef P. Debbins, Adam S. Fleisher, Marilyn Albert, Robert Green, George Bartzokis, Gary Glover, John Mugler, and Michael W. Weiner. The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. Journal of Mag- netic Resonance Imaging, 27(4):685–691, April 2008. ISSN 1053-1807, 1522-2586. doi: 10.1002/jmri.21049. URL https://onlinelibrary. wiley.com/doi/10.1002/jmri.21049. Jonghun Kim, Mansu Kim, and Hyunjin Park. Domain Aware Multi- task Pretraining of 3D Swin Transformer for T1-Weighted Brain MRI. In Minsu Cho, Ivan Laptev, Du Tran, Angela Yao, and Hongbin Zha, editors, Computer Vision – ACCV 2024, volume 15473, pages 121– 141. Springer Nature Singapore, Singapore, 2025. ISBN 978-981-96- 0900-0 978-981-96-0901-7. doi: 10.1007/978-981-96-0901-7_8. URL https://link.springer.com/10.1007/978-981-96-0901-7_8. Series Title: Lecture Notes in Computer Science. Arno Klein and Jason Tourville. 101 labeled brain images and a consistent human cortical labeling protocol. 6, 2012. ISSN 1662-453X. doi: 10. 3389/fnins.2012.00171. URL https://w.frontiersin.org/journals/ neuroscience/articles/10.3389/fnins.2012.00171/full. Daniel S. Marcus, Tracy H. Wang, Jamie Parker, John G. Csernansky, John C. Morris, and Randy L. Buckner. Open Access Series of Imaging Studies (OASIS): Cross-sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults. Journal of Cognitive Neuroscience, 19(9):1498–1507, September 2007. ISSN 0898-929X. doi: 10.1162/jocn.2007.19.9.1498. URL https://doi.org/10.1162/jocn. 2007.19.9.1498. Cheryl R. McCreary, Marina Salluzzi, Linda B. Andersen, David Gobbi, Louis Lauzon, Feryal Saad, Eric E. Smith, and Richard Frayne. Calgary Normative Study: design of a prospective longitudinal study to charac- terise potential quantitative MR biomarkers of neurodegeneration over the adult lifespan. BMJ Open, 10(8):e038120, August 2020. ISSN 2044-6055, 2044-6055. doi: 10.1136/bmjopen-2020-038120. URL https://bmjopen.bmj.com/content/10/8/e038120. Tasneem Nasser, Roberto Souza, and Naser El-Sheimy. Investigating Voxel-Level Brain Age Prediction as a Pretext Task for Brain MRI Segmentation. In James C. Gee, Daniel C. Alexander, Jaesung Hong, Juan Eugenio Iglesias, Carole H. Sudre, Archana Venkataraman, Polina Golland, Jong Hyo Kim, and Jinah Park, editors, Medical Image Com- puting and Computer Assisted Intervention – MICCAI 2025, pages 289– 299, Cham, 2026. Springer Nature Switzerland. ISBN 978-3-032- 05325-1. doi: 10.1007/978-3-032-05325-1_28. Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context Encoders: Feature Learning by Inpainting. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536–2544, Las Vegas, NV, USA, June 2016. IEEE. ISBN 978-1-4673-8851-1. doi: 10.1109/CVPR.2016.278. URL http: T. Nasser et al.: Preprint submitted to ElsevierPage 15 of 16 Investigating Pretext Tasks for Brain MR Segmentation //ieeexplore.ieee.org/document/7780647/. Martin Reuter, H. Diana Rosas, and Bruce Fischl. Highly accurate inverse consistent registration: A robust approach. 53(4):1181–1196. ISSN 1053-8119. doi: 10.1016/j.neuroimage.2010.07.020. URL https://w. sciencedirect.com/science/article/pii/S1053811910009717. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolu- tional Networks for Biomedical Image Segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi, editors, Medical Image Computing and Computer-Assisted Intervention – MIC- CAI 2015, pages 234–241, Cham, 2015. Springer International Publish- ing. ISBN 978-3-319-24574-4. doi: 10.1007/978-3-319-24574-4_28. Meredith A. Shafto, Lorraine K. Tyler, Marie Dixon, Jason R. Taylor, James B. Rowe, Rhodri Cusack, Andrew J. Calder, William D. Marslen- Wilson, John Duncan, Tim Dalgleish, Richard N. Henson, Carol Brayne, and Fiona E. Matthews. The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) study protocol: a cross-sectional, lifespan, multidisciplinary examination of healthy cognitive ageing. BMC Neu- rology, 14(1):1–25, December 2014. ISSN 1471-2377. doi: 10. 1186/s12883-014-0204-1. URL https://link.springer.com/article/ 10.1186/s12883-014-0204-1. Number: 1. Roberto Souza, Oeslle Lucena, Julia Garrafa, David Gobbi, Marina Saluzzi, Simone Appenzeller, Letícia Rittner, Richard Frayne, and Roberto Lotufo. An open, multi-vendor, multi-field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement. NeuroImage, 170:482–494, April 2018. ISSN 1053-8119. doi: 10. 1016/j.neuroimage.2017.08.021. URL https://w.sciencedirect.com/ science/article/pii/S1053811917306687. Catalin Stoean, Nebojsa Bacanin, Ruxandra Stoean, Leonard Ionescu, Cris- tian Alecsa, Mircea Hotoleanu, Miguel Atencia, and Gonzalo Joya. On Using Perceptual Loss within the U-Net Architecture for the Semantic Inpainting of Textile Artefacts with Traditional Motifs. In 2022 24th International Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC), pages 276–283, September 2022. doi: 10.1109/SYNASC57785.2022.00051. URL https://ieeexplore.ieee. org/abstract/document/10131036. Yucheng Tang, Dong Yang, Wenqi Li, Holger R. Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 20698–20708, New Orleans, LA, USA, June 2022. IEEE. ISBN 978-1-6654-6946-3. doi: 10.1109/CVPR52688.2022. 02007. URL https://ieeexplore.ieee.org/document/9879123/. W T Yuh, M R Crain, D J Loes, G M Greene, T J Ryals, and Y Sato. MR imaging of cerebral ischemia: findings in the first 24 hours. 12(4): 621–629, 1991. ISSN 0195-6108. URL https://pmc.ncbi.nlm.nih.gov/ articles/PMC8331595/. Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018. URL http://openaccess.thecvf.com/content_cvpr_2018/html/Zhang_The_ Unreasonable_Effectiveness_CVPR_2018_paper.html. Chen Zheng. Self-Supervised Pretext Tasks for Alzheimer’s Disease Classification using 3D Convolutional Neural Networks on Large-Scale Synthetic Neuroimaging Dataset, June 2024. URL http://arxiv.org/ abs/2406.14210. arXiv:2406.14210 [eess]. Xi-Nian Zuo, Jeffrey S. Anderson, Pierre Bellec, Rasmus M. Birn, Bharat B. Biswal, Janusch Blautzik, John C. S. Breitner, Randy L. Buckner, Vince D. Calhoun, F. Xavier Castellanos, Antao Chen, Bing Chen, Jiangtao Chen, Xu Chen, Stanley J. Colcombe, William Courtney, R. Cameron Craddock, Adriana Di Martino, Hao-Ming Dong, Xi- aolan Fu, Qiyong Gong, Krzysztof J. Gorgolewski, Ying Han, Ye He, Yong He, Erica Ho, Avram Holmes, Xiao-Hui Hou, Jeremy Huckins, Tianzi Jiang, Yi Jiang, William Kelley, Clare Kelly, Margaret King, Stephen M. LaConte, Janet E. Lainhart, Xu Lei, Hui-Jie Li, Kaiming Li, Kuncheng Li, Qixiang Lin, Dongqiang Liu, Jia Liu, Xun Liu, Yijun Liu, Guangming Lu, Jie Lu, Beatriz Luna, Jing Luo, Daniel Lurie, Ying Mao, Daniel S. Margulies, Andrew R. Mayer, Thomas Meindl, Mary E. Meyerand, Weizhi Nan, Jared A. Nielsen, David O’Connor, David Paulsen, Vivek Prabhakaran, Zhigang Qi, Jiang Qiu, Chunhong Shao, Zarrar Shehzad, Weijun Tang, Arno Villringer, Huiling Wang, Kai Wang, Dongtao Wei, Gao-Xia Wei, Xu-Chu Weng, Xuehai Wu, Ting Xu, Ning Yang, Zhi Yang, Yu-Feng Zang, Lei Zhang, Qinglin Zhang, Zhe Zhang, Zhiqiang Zhang, Ke Zhao, Zonglei Zhen, Yuan Zhou, Xing-Ting Zhu, and Michael P. Milham. An open science resource for establishing reliability and reproducibility in functional connectomics. Scientific Data, 1(1):140049, December 2014. ISSN 2052-4463. doi: 10.1038/ sdata.2014.49. URL https://w.nature.com/articles/sdata201449. T. Nasser et al.: Preprint submitted to ElsevierPage 16 of 16