Paper deep dive
Multiple Myeloma Lesion Segmentation on Whole-Body Diffusion-Weighted Imaging via Efficient Anatomical Anticipation and Multimodal Confirmation
Mengmeng Zhang, Shengqian Huang, Junde Zhou, Xiaoping Wu, Hao Luog, Jing Wanga, Yicheng Sun, Jiao Li, Haibo Zhang, Sheng Xie, Fan Wangg, Qin Wangc, Huadan Xue, Yisheng Lv, Fei-yue Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 9/9/2026, 6:00:16 AM
Summary
The paper proposes a two-stage framework for automated multiple myeloma (MM) lesion segmentation on whole-body diffusion-weighted imaging (WB-DWI). Stage 1 generates bone region-of-interest (ROI) maps from Apparent Diffusion Coefficient (ADC) images using a 3D nnUNet ensemble, leveraging lesion labels from diffuse cases as surrogate supervision to avoid manual bone annotation. Stage 2 employs the Anatomy-guided Multimodal U-Net (AMU-Net), which integrates ADC and WB-DWI via encoder-level ROI-gated cross-attention for structural supplementation and a decoder-level asymmetric gate for lesion confirmation. The method achieves a mean Dice score of 76.2%, outperforming existing baselines.
Entities (8)
Relation Signals (7)
AMU-Net → achievesperformance → 76.2% Dice score
confidence 98% · It achieves the best overall performance among the evaluated methods, with a mean Dice score of 76.2%.
3D nnUNet → generates → Bone ROI
confidence 95% · Stage I generates bone ROI maps from ADC images using a 3D nnUNet ensemble.
Multiple Myeloma → isassessedby → WB-DWI
confidence 95% · Whole-body diffusion-weighted imaging (WB-DWI) is widely used for multiple myeloma (MM) assessment
AMU-Net → uses → Bone ROI
confidence 95% · Stage II performs lesion segmentation on WB-DWI under explicit bone ROI guidance from Stage I.
AMU-Net → uses → WB-DWI
confidence 95% · Stage II performs lesion segmentation on WB-DWI under explicit bone ROI guidance from Stage I. AMU-Net, is built on a 3D U-Net backbone with DWI as the primary input
3D nnUNet → uses → ADC
confidence 95% · Stage I: Bone ROI Generation (ADC àBone ROI)... train... ADC
AMU-Net → uses → ADC
confidence 95% · AMU-Net incorporates ADC information in two complementary ways... ADC as an auxiliary modality.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Whole-body diffusion-weighted imaging (WB-DWI) is widely used for multiple myeloma (MM) assessment, yet automated lesion segmentation remains challenging due to limited anatomical delineation and the low specificity of marrow hyperintensity. Existing studies have introduced bone region-of-interest (ROI) information and apparent diffusion coefficient (ADC) maps to mitigate these ambiguities, but practical limitations remain. Bone ROI construction often relies on costly manual annotation, image registration, or dedicated bone models, while ADC is usually incorporated only through simple channel fusion, limiting its ability to provide complementary structural and lesion-discriminative cues. To address these limitations, we propose a two-stage framework for MM lesion segmentation on WB-DWI. In the first stage, we train a bone ROI generation model from ADC images without dedicated bone labels, providing an efficient and practical anatomical prior for lesion analysis. In the second stage, we propose Anatomy-guided Multimodal U-Net (AMU-Net), which leverages ADC in a manner consistent with clinical lesion assessment rather than treating it as a generic auxiliary modality. Extensive experiments demonstrate the effectiveness and practicality of the proposed method. It achieves the best overall performance among the evaluated methods, with a mean Dice score of 76.2%.
Tags
Links
- Source: https://arxiv.org/abs/2609.06165v1
- Canonical: https://arxiv.org/abs/2609.06165v1
Trouble viewing inline? Open PDF directly →
Full Text
48,717 characters extracted from source content.
Expand or collapse full text
Multiple Myeloma Lesion Segmentation on Whole-Body Diffusion-Weighted Imaging via Efficient Anatomical Anticipation and Multimodal Confirmation Mengmeng Zhang a,b , Shengqian Huang c,1 , Junde Zhou c , Xiaoping Wu g , Hao Luo g,h , Jing Wang a,f , Yicheng Sun a,b , Jiao Li c,d , Haibo Zhang e , Sheng Xie e , Fan Wang g , Qin Wang c,∗ , Huadan Xue c,∗ , Yisheng Lv a,b,∗ and Fei-yue Wang a,∗ a the State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, Beijing, China b the School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 101408, Beijing, China c the Department of Radiology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, 100730, Beijing, China d the Department of Medical Imaging, Peking University Shenzhen Hospital, Shenzhen, 518036, Guangdong, China e the Department of Radiology, China-Japan Friendship Hospital, Beijing, 100029, Beijing, China f the Faculty of Innovation Engineering, Macau University of Science and Technology, Macao, 999078, Macao, China g the DAMO Academy, Alibaba Group, Hangzhou, Zhejiang, China h the Hupan Lab, Zhejiang Province Hangzhou, Zhejiang, China A R T I C L E I N F O Keywords: Multiple myeloma whole-body DWI lesion segmentation region of interest(ROI) apparent diffusion coefficient(ADC) anatomy-constrained fusion A B S T R A C T Whole-body diffusion-weighted imaging (WB-DWI) is widely used for multiple myeloma (M) assessment, yet automated lesion segmentation remains challenging due to limited anatomical delineation and the low specificity of marrow hyperintensity. Existing studies have introduced bone region-of-interest (ROI) information and apparent diffusion coefficient (ADC) maps to mitigate these ambiguities, but practical limitations remain. Bone ROI construction often relies on costly manual annotation, image registration, or dedicated bone models, while ADC is usually incorporated only through simple channel fusion, limiting its ability to provide complementary structural and lesion- discriminative cues. To address these limitations, we propose a two-stage framework for M lesion segmentation on WB-DWI. In the first stage, we train a bone ROI generation model from ADC images without dedicated bone labels, providing an efficient and practical anatomical prior for lesion analysis. In the second stage, we propose Anatomy-guided Multimodal U-Net (AMU-Net), which leverages ADC in a manner consistent with clinical lesion assessment rather than treating it as a generic auxiliary modality. Extensive experiments demonstrate the effectiveness and practicality of the proposed method. It achieves the best overall performance among the evaluated methods, with a mean Dice score of 76.2%. 1. Introduction Whole-body diffusion-weighted imaging (WB-DWI) is widely used for diagnosis and response assessment in mul- tiple myeloma (M) because it provides whole-body cover- age without ionizing radiation or contrast administration [26, 18]. Automated lesion segmentation on WB-DWI would be clinically valuable, but remains challenging. Although ∗ Corresponding author zhangmengmeng2022@ia.ac.cn (M. Zhang); sqhuang18@163.com (S. Huang); zhoujunde@pumch.cn (J. Zhou); xpwu95@163.com (X. Wu); michuan.lh@alibaba-inc.com (H. Luo); wangjing2014@ia.ac.cn (J. Wang); sunyicheng2025@ia.ac.cn (Y. Sun); ljlyyoung@163.com (J. Li); zhb\_hello@163.com (H. Zhang); xs\_mri@126.com (S. Xie); fan.w@alibaba-inc.com (F. Wang); wangqin@pumch.cn (Q. Wang); xuehd@pumch.cn (H. Xue); yisheng.lv@ia.ac.cn (Y. Lv); feiyue.wang@ia.ac.cn (F. Wang) ORCID(s): 0009-0003-8210-8292 (M. Zhang); 0000-0003-3257-6082 (S. Huang); 0009-0005-7223-6316 (J. Zhou); 0000-0002-3067-4626 (X. Wu); 0000-0002-6405-4011 (H. Luo); 0000-0002-4419-9352 (J. Wang); 0009-0002-4560-5196 (Y. Sun); 0000-0002-2574-276X (J. Li); 0000-0001-5327-0395 (H. Zhang); 0000-0003-3280-9632 (S. Xie); 0000-0002-2970-2642 (Q. Wang); 0000-0002-4278-2165 (H. Xue); 0000-0002-0508-1298 (Y. Lv) 1 This author contributed equally to this work as M.Zhang WB-DWI highlights diffusion-restricted lesions, it provides limited anatomical delineation of the skeleton, making it difficult to distinguish marrow lesions from adjacent soft tissue or other extraskeletal structures. In addition, marrow hyperintensity is not specific to M and may also arise from benign marrow hyperplasia, inflammation, edema, and other nonmalignant processes. In clinical reading, suspicious hyperintensity on WB- DWI is therefore not interpreted in isolation. Radiologists rely on both anatomical context and ADC maps derived from the same diffusion acquisition. Anatomical knowledge helps determine whether suspicious signal is plausibly located within marrow-containing bone, while ADC provides com- plementary tissue characterization for lesion assessment [24, 28, 2]. Motivated by this reading strategy, previous studies have introduced bone ROI information and ADC into M lesion segmentation pipelines [3, 10]. However, important limitations remain. Bone ROI construction often relies on manual annotation, atlas-based registration, or dedicated bone models [31, 5, 6], all of which increase cost and complexity. Meanwhile, ADC is typically incorporated only Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 1 of 11 arXiv:2609.06165v1 [cs.CV] 5 Sep 2026 through simple input-level concatenation, without explicitly modeling its distinct roles in structural supplementation and lesion confirmation. To address these limitations, we propose a two-stage framework for M lesion segmentation on WB-DWI. In the first stage, we train a bone ROI generation model from in- herently co-registered ADC images without dedicated bone labels, producing an anatomical prior for downstream lesion segmentation. In the second stage, we propose Anatomy- guided Multimodal U-Net (AMU-Net), which leverages ADC in a manner consistent with clinical lesion assess- ment. Specifically, ADC is used to supplement anatomical structure at the encoder and to support lesion confirmation at the decoder under explicit ROI guidance. Our main contributions can be summarized as follows: • We develop a bone ROI generation strategy that learns skeletal priors from ADC images without dedicated bone labels, avoiding additional bone annotation and registration. • We propose Anatomy-guided Multimodal U-Net (AMU- Net) for M lesion segmentation on WB-DWI, which uses ADC for both anatomy-aware structural supple- mentation and lesion confirmation under ROI guid- ance. • We validate the proposed design through system- atic experiments and ablation studies, showing that annotation-efficient ROI generation and clinically mo- tivated ADC utilization lead to improved M lesion segmentation performance. To facilitate community use, we will publicly release the code repository and trained model weights upon acceptance at https://github.com/M-MZhang/M_segmentation.git. 2. Related work 2.1. Bone ROI Acquisition and Anatomy-Aware Segmentation Previous studies have recognized that bone or bone mar- row region-of-interest (ROI) information can serve as an ef- fective anatomical prior for multiple myeloma (M) lesion segmentation, as it helps suppress false positives and con- strains predictions to plausible marrow-bearing regions [33]. This has motivated a line of work on automatic bone or bone marrow ROI acquisition [19, 27, 22, 30, 5]. However, most such methods have been developed on modalities with clearer structural depiction, such as CT or anatomical MRI, rather than on WB-DWI, which is more commonly used for lesion assessment in M [17, 3]. On WB-DWI, the weak delineation of skeletal and marrow boundaries makes both manual annotation and direct structure learning substan- tially more challenging. Consequently, methods developed on other modalities are not readily applicable to WB-DWI- based lesion segmentation. The relatively few attempts to obtain skeletal ROIs directly on WB-DWI have mainly relied on atlas- or registration-based localization [5], which is Two-stage M Lesion Segmentation Framework Clinical Reading Strategy DWI Skeletal prior Bone Anticipation ADC Confirm Bone ROI Generation Bone ROI Generation Model Lesion Label AMU-Net ADC Incorporation Prior Work Manual annotation Atlas DWI ADC Simple channel fusion Labor-intensive Inaccurate Incomplete information utilization Stag 1 ADC Lesion Label train Stag 2 ROI-gated Cross- Attention Asymmetric Gate Bone Label-Free Straightforward Clinically Consistent Effective Figure 1: Motivation of the proposed framework. In clinical reading, radiologists first anticipate anatomically plausible marrow regions as skeletal priors and then jointly interpret WB- DWI and ADC to confirm M lesions. Prior work addresses these cues only partially, relying on manual or atlas-based bone ROI generation and simple ADC channel fusion. Inspired by this strategy, our method adopts a two-stage design: Stage 1 learns a bone ROI generation model from ADC without dedicated bone labels, and Stage 2 uses AMU-Net for anatomy-guided multimodal lesion segmentation. often time-consuming and sensitive to variations in patient positioning, field-of-view, and image appearance, thereby limiting its robustness in practice. 2.2. ADC-Guided Multimodal Fusion for M Lesion Segmentation Apparent diffusion coefficient (ADC) maps are a natural auxiliary input for multiple myeloma (M) lesion seg- mentation because they are derived from the same WB- DWI acquisition and provide quantitative diffusion-related information [14, 4, 32]. Since ADC maps are inherently aligned with WB-DWI and reflect tissue cellularity, they can provide complementary evidence for distinguishing true pathological diffusion restriction from normal marrow het- erogeneity. Existing M methods that use ADC typically adopt simple input-level fusion by concatenating ADC with WB-DWI as an additional channel [1]. While effective as a straightforward baseline, this strategy treats ADC as a generic secondary modality and does not explicitly cap- ture its task-specific role in M, where it may assist both lesion confirmation and structural delineation within the marrow. More advanced multimodal fusion methods, such as feature-wise modulation, gated fusion, and attention-based interaction, have been explored in broader medical image Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 2 of 11 Stage 1: Bone ROI Generation (ADC àBone ROI) Stage 2: AMU-Net (Anatomy-guided Multimodal U-Net) Cross- Attention MLP Gate Cross- Attention MLP Gate Cross- Attention MLP Gate ROI-gated Cross-Attention 3 levels Agree Disagree tanh sigmoid Asymmetric Gate ! ! " " # # ! $ #$% % % & & 3D U-Net ' " Conv Block + Down sample Conv Block + Up sample ' !=!×(++-./0) Minus sign Add signContrapositionalmultiplication ! #&"′ ! $()#&"′ nnUnet Bone Segmentation Model Input ADCFold 0 Fold 1 Fold 2Fold 3 Fold 4 Generated Bone ROI mean Generated ROI used in stage 2 ADC WB-DWI Figure 2: Overview of the proposed two-stage framework for M lesion segmentation on WB-DWI. Stage I generates bone ROI maps from ADC images using a 3D nnUNet ensemble. Stage I uses AMU-Net to perform ROI-guided lesion segmentation from WB-DWI with ADC assistance, via encoder-level ROI-gated cross-attention and a decoder-level asymmetric gate. segmentation settings [8, 20, 35, 36], but they are mainly evaluated on modality pairs such as multiparametric MRI or CT/PET. Their suitability for M lesion segmentation on WB-DWI therefore remains unclear. 3. Method 3.1. Overview The method is organized as a two-stage framework for multiple myeloma (M) lesion segmentation on whole- body diffusion-weighted imaging (WB-DWI). Bone ROI maps are first generated from ADC images, after which lesion segmentation is performed by ADC-guided Multi- modal U-Net (AMU-Net). The overall architecture com- prises three key components: (i) annotation-efficient bone ROI generation, (i) encoder-level ROI-gated ADC–DWI interaction for structural supplementation, and (i) decoder- level asymmetric gate for lesion confirmation, as shown in Fig. 2. 3.2. Stage I: Annotation-Efficient Bone ROI Generation To achieve annotation-efficient bone ROI generation, we exploit a subset of diffuse M cases in which lesion involvement spans most of the marrow cavity. In these cases, the available lesion annotations can serve as surrogate supervision for skeletal marrow extent, thereby avoiding ad- ditional manual bone annotation. Combined with the clearer structural delineation provided by ADC compared with WB- DWI, this enables learning of a task-specific ROI generator for downstream lesion segmentation. Specifically, let 퐴 ∈ℝ 퐻×푊 ×퐷 denote the ADC vol- ume, and let 푅 ∈ [0, 1] 퐻×푊 ×퐷 denote the predicted bone ROI probability map. We train a 3D nnUNet ensemble on ADC using lesion labels from diffuse M cases as pseudo- marrow supervision. The goal of this model is not precise anatomical bone segmentation, but rather generation of a soft ROI prior with sufficient skeletal coverage for subse- quent lesion segmentation. Each ensemble member 푚 = 1, ...,푀 predicts a probability map 푅 (푚) , and the final ROI Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 3 of 11 is obtained by voxel-wise mean aggregation: 푅(퐯) = 1 푀 푀 ∑ 푚=1 푅 (푚) (퐯),(1) where 퐯 indexes voxel location. This aggregation reduces prediction variance across ensemble members and produces a more stable soft anatomical prior for Stage I. Because ADC is acquired in the same diffusion sequence as DWI, the ROI map is inherently aligned with the lesion segmentation input and can be used directly without addi- tional registration. 3.3. Stage I: AMU-Net Stage I performs lesion segmentation on WB-DWI un- der explicit bone ROI guidance from Stage I. AMU-Net, is built on a 3D U-Net backbone with DWI as the primary input for lesion representation and ADC as an auxiliary modal- ity. Guided by the clinical practice of jointly interpreting DWI and ADC, AMU-Net incorporates ADC information in two complementary ways. At the encoder level, ROI-gated cross-attention is used to inject structural cues from ADC into DWI features, compensating for the limited anatomical delineation of WB-DWI. At the decoder level, a decoder- level asymmetric gate modulates the final lesion embedding according to DWI–ADC concordance for lesion confirma- tion. Together, these designs enable ROI- and ADC-aware lesion segmentation while preserving the lesion-sensitive representation learned from DWI. 3.3.1. Encoder-level ROI-gated cross-attention WB-DWI enhances lesion conspicuity by suppressing background signal, but this also weakens anatomical depic- tion of the skeleton. To supplement the missing structural information, we introduce ROI-gated ADC–DWI cross- attention at multiple encoder levels. Let 퐹 푑 푙 and 퐹 푎 푙 denote the DWI and ADC feature maps at encoder level 푙, respectively, and let 푅 푙 denote the Stage- I ROI map resized to the same spatial resolution. For each selected level, the two feature maps are concatenated and passed through a lightweight gating function: 퐺 푙 = 휎 ( 휙 푙 ([퐹 푑 푙 ,퐹 푎 푙 ]) ) ⊙ 푅 푙 (2) where [⋅,⋅ ] denotes channel-wise concatenation, 휙 푙 (⋅) is a two-layer 1 × 1 × 1 convolutional projection with ReLU activation, 휎(⋅) denotes the sigmoid function, and ⊙ denotes element-wise multiplication. The DWI feature is then modulated as ̃ 퐹 푑 푙 = 퐹 푑 푙 ⊙ (1 + 퐺 푙 )(3) This design allows ADC to supplement DWI features with clearer structural cues while preserving DWI as the pri- mary lesion-sensitive representation. The ROI gate further constrains cross-modal interaction to likely skeletal regions, reducing interference from extra-skeletal tissue. We apply this interaction at the first three encoder levels, as shallow encoder features preserve richer spatial and anatomical de- tail, making them more suitable for ADC-based structural supplementation than deeper semantic features. 3.3.2. Decoder-level asymmetric gate Within the marrow cavity, suspicious DWI hyperinten- sity is not specific to M lesions and is often interpreted together with ADC in clinical reading. To further improve lesion confirmation, we introduce a decoder-level asymmet- ric gate that modulates the final DWI embedding according to DWI–ADC concordance and discrepancy. Let 퐸 ∈ℝ 퐶×퐻×푊 ×퐷 denote the final full-resolution DWI decoder embedding, and let 퐹 푎 ∈ℝ 퐶 푎 ×퐻×푊 ×퐷 denote the corresponding full-resolution ADC feature map. We first inject the explicit skeletal prior by ROI-based amplification: 퐸 푟 = 퐸 ⊙ (1 + 푅)(4) where the ROI map 푅 is broadcast along the channel dimen- sion. Next, point-wise linear projections map the DWI and ADC features into a shared latent space: 푄 = 푊 푞 (퐸 푟 ) 퐾 = 푊 푘 (퐹 푎 )(5) where 푊 푞 and 푊 푘 are learnable 1 × 1 × 1 convolutions. The agree branch captures joint DWI–ADC evidence: 퐴 agree = tanh ( 휓 푎 (푄 + 퐾) ) (6) where 휓 푎 (⋅) is a 1 × 1 × 1 convolution producing a single- channel response map. The tanh activation allows this branch to either enhance or suppress the embedding depend- ing on the local joint DWI–ADC pattern. The disagree branch captures discrepant evidence be- tween DWI and ADC: 퐴 disagree = 휎 ( 휓 푑 (푄 − 퐾) ) (7) where 휓 푑 (⋅) is another 1 × 1 × 1 convolution and 휎(⋅) ∈ [0, 1]. This branch acts as a one-sided suppressive term for potentially misleading cross-modality discrepancy. The combined attention map is defined as 퐴 = 퐴 agree − 퐴 disagree (8) To reduce the influence of unreliable ADC responses, we further introduce a simple ADC confidence term 퐶 adc . Although ADC values are expected to be non-negative by definition, voxel-wise ADC measurements may still be af- fected by the long WB-DWI acquisition process and related noise, distortion, and motion artifacts. We therefore define 퐶 adc based on whether the local ADC value is positive: 퐶 adc (푥) = 1, ADC(푥) > 0 0, otherwise (9) and use it to restrict the attention modulation to locations with valid ADC support: 퐴 ′ = 퐴 ⊙ 퐶 adc (10) The refined decoder embedding is then computed as ̃ 퐸 = 퐸 푟 + 퐴 ′ ⊙ 퐸 푟 (11) Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 4 of 11 Center A (n=67) Center B (n=29) 40 50 60 70 80 Age (years) (a) Age Distribution by Center Center A (n=67) Center B (n=29) 0 20 40 60 80 Count 54% 55% 46% 45% (b) Sex Distribution by Center Male Female Center A (n=67) Center B (n=29) 0 20 40 60 80 100 Percentage (%) Focal 27 (40%) Focal 3 (10%) Diffuse 39 (58%) Diffuse 19 (66%) Normal 7 (24%) (c) Subtype Distribution by Center Focal Diffuse Normal Ctr A Focal Ctr A Diffuse Ctr B Focal Ctr B Diffuse 40 50 60 70 80 Age (years) (d) Age by Subtype × Center Figure 3: Demographic and clinical characteristics of the internal and external cohorts: (a) age distribution by center, (b) sex distribution by center, (c) subtype distribution by center, and (d) age distribution stratified by subtype and center. Blue and green denote the internal and external cohorts, respectively. Violin plots show density distributions with embedded boxplots. FocalDiffuse 0 500 1000 1500 2000 2500 Lesion Burden (mL) (a) Lesion Burden FocalDiffuse 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Lesion Sparsity r ( ×10 3 ) (b) Lesion Sparsity (r = n lesion / n body ) LesionMarrowBackground 0 50 100 150 200 250 DWI Intensity (raw) (c) DWI by Tissue Type LesionMarrowBackground 0.0 0.5 1.0 1.5 2.0 ADC ( ×10 3 m²/s) (d) ADC by Tissue Type Figure 4: Task-related data characteristics of the internal cohort: (a) lesion burden, (b) lesion sparsity, (c) DWI signal distributions by tissue type, and (d) ADC distributions by tissue type. Green and blue denote focal and diffuse cases in (a) and (b), while colors in (c) and (d) indicate lesion, marrow, and background tissues. Violin plots show density distributions with embedded boxplots, and dots represent individual observations. This asymmetric design distinguishes concordance from discrepancy in DWI–ADC interaction, while the confidence term limits ADC-guided modulation to locations with valid ADC support. The refined embedding ̃ 퐸 is finally fed to the classification head to produce lesion logits. 3.4. Training Objectives The two stages are trained separately. Stage I follows the standard nnUNet training setup for ROI prediction. After training, the Stage-I model is fixed and used to generate ROI probability maps for Stage I. For lesion segmentation, let 푌 and 푌 ∗ denote the pre- dicted and reference lesion masks, respectively. Stage I is optimized using a combination of Dice loss and binary cross- entropy (BCE) loss: seg = 휆 1 Dice (푌 ,푌 ∗ ) + 휆 2 BCE (푌 ,푌 ∗ )(12) where 휆 1 and 휆 2 are balancing weights. Unless otherwise specified, both are set to 1. 4. Experiments 4.1. Dataset and Data Analysis 4.1.1. Cohort Description This study used two whole-body MRI cohorts collected from different institutions. The internal cohort, acquired at Peking Union Medical College Hospital, comprised 67 cases. Among them, 47 cases were used for model devel- opment and were further evaluated under a 5-fold cross- validation protocol, while the remaining 20 cases were re- served as an independent internal validation set. The exter- nal cohort, acquired at China-Japan Friendship Hospital, comprised 29 cases and was reserved for independent exter- nal validation. The distributions of M subtypes, age, and sex in the two cohorts are summarized in Fig. 3. 4.1.2. Lesion Distribution Characteristics To characterize the complexity of whole-body M le- sion segmentation, we quantified lesion burden and lesion sparsity on the internal cohort, as shown in Fig. 4(a)–(b). Lesion burden was defined as the total lesion volume per patient, measured in milliliters from manual voxel-wise an- notations. Lesion sparsity was defined as the ratio of lesion- positive voxels to body voxels, i.e., 푟 = 푛 lesion ∕푛 body , reflecting how sparsely lesions are distributed within the whole-body field of view. Body voxels were obtained from the foreground body region after preprocessing, and lesion voxels were counted from the corresponding expert anno- tations. These statistics show that diffuse cases generally exhibited substantially larger lesion burden and higher lesion occupancy than focal cases. At the same time, the lesion- positive ratio remained extremely low overall, indicating severe class imbalance and sparse target distribution. Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 5 of 11 Table 1 Internal comparison with baseline models under the same 5-fold evaluation protocol. Best results are shown in bold, and second- best results are shown in blue. ModelTypeParams Dice (%)↑ IoU (%)↑ HD95↓ Precision (%)↑ Recall (%)↑ nnU-Net[16] 3D CNN ∼30M 73.13 ± 1.5 63.24 ± 1.3 11.8 ± 1.5ퟖퟎ.ퟎ76.3 VISTA3D[13]∼30M 72.40 ± 0.4 61.43 ± 0.7 12.6 ± 0.573.0ퟖퟐ.ퟑ SegResNet[21]∼19M 67.04 ± 0.7 56.00 ± 0.7 16.3 ± 0.867.376.7 U-Net[25]∼5M 64.92 ± 1.4 52.93 ± 1.5 16.5 ± 0.568.373.5 DynUNet[7]∼5M 63.81 ± 1.7 51.93 ± 1.9 16.7 ± 1.365.874.0 AttentionUNet[23]∼5M 63.29 ± 0.5 51.45 ± 0.8 16.2 ± 0.765.576.1 SwinUNETR[11] 3D Transformer ∼62M 63.72 ± 0.4 52.11 ± 0.6 19.4 ± 0.564.674.8 UNETR[12]∼93M 57.32 ± 1.1 45.09 ± 1.1 24.3 ± 1.053.278.1 MA-Net[9] 2D CNN ∼28M 72.75 ± 1.4 61.96 ± 1.6 11.9 ± 1.471.482.1 UNeXt[29]∼1.5M 71.67 ± 0.2 60.12 ± 0.3 12.0 ± 0.770.681.0 MultiResUNet[15]∼7M 68.32 ± 1.5 56.02 ± 1.9 13.1 ± 1.365.579.0 MMNet[34]∼1.5M 68.46 ± 3.3 56.70 ± 4.3 15.2 ± 3.068.177.9 AMU-Net(Ours)–∼31M ퟕퟔ.ퟐퟒ ± ퟎ.ퟓ ퟔퟓ.ퟗퟑ ± ퟏ.ퟎ ퟏퟎ.ퟏ ± ퟎ.ퟕ78.382.1 Table 2 Zero-shot external validation on an independent cohort. Models trained using the same 5-fold protocol were directly tested on the external dataset without any fine-tuning or adaptation. Best results are shown in bold, and second-best results are shown in blue. ModelTypeParams Dice (%)↑ IoU (%)↑HD95↓Precision (%)↑ Recall (%)↑ nnU-Net[16] 3D CNN ∼30M 28.70 ± 1.3 18.59 ± 1.1 70.29 ± 0.8232.3135.16 VISTA3D[13]∼30M 30.83 ± 1.1 21.26 ± 0.9 61.49 ± 0.7926.4448.78 SegResNet[21]∼19M 40.65 ± 0.9 29.96 ± 0.8 53.41 ± 5.0244.7944.93 U-Net[25]∼5M39.64 ± 2.7 28.97 ± 2.5 63.98 ± 3.0236.9953.28 DynUNet[7]∼5M39.42 ± 2.4 28.55 ± 2.2 62.20 ± 4.4635.4356.46 AttentionUNet[23]∼5M35.99 ± 1.4 25.58 ± 1.2 69.05 ± 6.7232.46ퟓퟕ.ퟖퟎ SwinUNETR[11] 3D Transformer ∼62M 35.43 ± 1.8 24.99 ± 1.7 66.33 ± 2.8732.2650.30 UNETR[12]∼93M 30.77 ± 0.5 20.74 ± 0.4 68.39 ± 4.5125.0654.22 MA-Net[9] 2D CNN ∼28M 37.26 ± 1.7 27.62 ± 1.5 50.47 ± 5.1039.7941.26 UNeXt[29]∼1.5M 38.11 ± 2.5 28.00 ± 2.1 56.03 ± 1.4835.8851.10 MultiResUNet[15]∼7M37.86 ± 2.6 27.28 ± 2.1 46.34 ± 7.3444.6937.77 MMNet[34]∼1.5M 30.66 ± 2.9 22.11 ± 2.4 57.56 ± 1.2428.4849.05 AMU-Net(Ours)–∼31M 43.43 ± 0.9 33.04 ± 0.9 59.93 ± 3.3148.0047.43 4.1.3. DWI and ADC Signal Characteristics We further analyzed modality-specific tissue signals us- ing 247 radiologist-annotated points from 8 selected cases, covering lesion, marrow, and background regions (Fig. 4(c)– (d)). For each annotated point, the corresponding DWI and ADC values were sampled from the original images to examine tissue-level signal distributions. The resulting plots show distinct yet partially overlapping signal patterns across tissue types in both modalities, indicating that DWI and ADC provide complementary but individually insufficient information for lesion identification. Overall, these analyses highlight several key challenges of M lesion segmentation on WB-DWI, including large inter-patient variation in lesion extent, extremely sparse lesion occupancy, and heteroge- neous multimodal signal characteristics. 4.2. Experimental Setup To evaluate the effectiveness and generalizability of the proposed method, all primary experiments were performed using a patient-level 5-fold cross-validation strategy. Unlike sample-level splitting, patient-level partitioning ensures that all images or records associated with the same patient are assigned to the same fold, thereby preventing information leakage between the training and test sets. The dataset was randomly divided into five mutually exclusive folds at the patient level. For each iteration, four folds were used for training and internal validation, and the remaining fold was Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 6 of 11 0.00.20.40.60.81.01.2 ADC (×10 3 m 2 /s) 0 5 10 15 20 25 30 Number of points (a) Pooled across all cases Normal marrow Lesion 0.00.20.40.60.81.01.2 ADC (×10 3 m 2 /s) Case 1-1 Case 11-3 Case 3-31 Case 3-37 Case 4-11 Case 4-2 Case 4-30 Case 4-61 (b) Per-case (mean ± range) LesionNormal marrowMean shift Figure 5: ADC distributions of lesion and normal marrow from pooled and per-case perspectives. (a) Pooled across all cases, lesion and normal marrow ADC values substantially overlap, suggesting that absolute ADC values alone are insufficient for disease assessment. (b) Per-case mean and range show that lesion-related abnormalities are more meaningfully assessed relative to the patient’s own marrow background. This supports the role of bone ROI guidance in constraining DWI–ADC interpretation to the skeletal marrow space. used for testing. This process was repeated five times, with each fold used once as the test set. All compared methods were evaluated under the same cross-validation protocol and data splits to ensure a fair comparison. The final quantitative results were reported as the mean and standard deviation across the five folds, providing an estimate of both the overall performance and the variability of each method. In addition to internal cross- validation, an independent external cohort was used to fur- ther evaluate the generalization ability of the model. The external cohort was excluded from all training and model selection procedures and was used solely for independent testing. Segmentation performance was comprehensively eval- uated using Dice, intersection over union (IoU), 95th per- centile Hausdorff distance (HD95), precision, and recall. These metrics respectively quantify region overlap, bound- ary accuracy, and the trade-off between false positive and false negative predictions. Results are reported as the mean and standard deviation across the five folds unless otherwise noted. Table 3 Ablation results of Stage-I components in AMU-Net. Enc- XAttn: encoder-level ROI-gated cross-attention; Dec-Gate: decoder-level asymmetric gate. VariantDice (%)↑ IoU (%)↑ HD95↓ Structural ablation Full76.24±0.5 65.93±1.0 10.1±0.7 w/o Enc-XAttn 73.45±0.8 63.52±1.1 12.5±0.5 w/o Dec-Gate74.87±0.7 64.74±0.9 11.0±0.5 w/o Both73.13±1.5 63.24±1.3 11.8±1.5 ROI guidance ablation w/o Enc-ROI75.80±0.4 65.55±0.5 11.3±0.5 w/o Dec-ROI75.10±0.6 64.94±0.8 11.9±0.5 4.3. Main Results 4.3.1. ROI Coverage Analysis We first examined whether the generated bone ROI could provide meaningful guidance for lesion segmenta- tion. As shown in Fig. 6, the ADC-based ROI generator produced anatomically coherent pseudo-skeletal structures even in cases with sparse lesion annotations, indicating that it learned a structural bone prior rather than merely fitting lesion labels. Despite the sparsity of the annotations, the generated ROI still covered most lesion regions in both focal and diffuse cases, supporting its role as an effective anatomical guide for downstream segmentation. 4.3.2. Internal Benchmark Results Table 1 summarizes the internal benchmark results. We compared AMU-Net against a diverse set of strong baselines, including representative 3D CNN, 3D transformer, and 2D CNN models, under the same 5-fold evaluation protocol. Among all evaluated methods, AMU-Net achieved the best overall performance, with the highest Dice (76.24±0.5) and IoU (65.93±1.0) and the lowest HD95 (10.1±0.7), indi- cating superior overlap accuracy and boundary delineation. Notably, nnU-Net was the strongest competing baseline, yet AMU-Net still improved Dice by 3.1% and further re- duced HD95. These results demonstrate that the proposed anatomy-guided and ADC-aware design yields consistent gains over a broad range of competitive segmentation archi- tectures. AMU-Net did not achieve the highest precision or recall individually, but this does not indicate weaker detection abil- ity. In practice, very high precision often reflects a conserva- tive prediction strategy that suppresses false positives at the cost of more missed lesions, whereas very high recall usually corresponds to a more aggressive strategy that captures more lesions but also introduces more false-positive predictions. By achieving the second-best precision and tied second- best recall, AMU-Net showed a better balance between lesion sensitivity and false-positive control. This trend is also consistent with the qualitative comparisons in Fig. 7, where AMU-Net showed more complete lesion coverage with fewer obvious misses. Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 7 of 11 Figure 6: Visualization of ADC-based bone ROI generation. Representative sparse/dense and focal/diffuse cases are shown to illustrate ROI generation across heterogeneous patients. In all cases, the predicted ROI forms anatomically coherent skeletal structures while covering a substantial portion of lesion regions. Although trained with lesion-derived supervision, the ROI generator does not simply fit sparse lesion labels, but instead learns stable bone-related structures for downstream segmentation. Figure 7: Visual comparison of lesion segmentation results. Representative diffuse and focal cases show that AMU-Net achieves more complete lesion coverage and fewer obvious misses than the compared methods. Zoomed-in views are provided for the boxed regions. Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 8 of 11 Figure 8: Qualitative comparison between a registration pipeline and the proposed direct bone ROI generation method.The registration-based method often shows spatial mismatch after mapping Dixon-derived bone masks into DWI space, whereas our method produces bone ROIs that better align with lesion-bearing skeletal regions. 4.3.3. External Validation Results To further evaluate the generalization capability of AMU-Net, we conducted a zero-shot external validation on the independent cohort collected from Tianjin Institute of Hematology Hospital, without any fine-tuning or domain adaptation. As shown in Table 2, the overall performance decreased for all methods compared with the internal eval- uation, reflecting the substantial distribution shift between institutions. In particular, the external cohort contains a considerably larger proportion of normal cases, whereas only one normal case was available during training. Such a mismatch makes lesion localization substantially more challenging and lowers the overall quantitative metrics across all competing methods. Despite this challenging setting, AMU-Net consistently achieved the best overall performance, outperforming all baseline methods in Dice and IoU while maintaining com- petitive Precision and Recall. Notably, nnU-Net, which ranked second on the internal validation, exhibited a much larger performance degradation on the external cohort, whereas AMU-Net retained a clear performance margin. These results suggest that the proposed anatomy-aware gat- ing mechanism, designed according to clinical prior knowl- edge and lesion distribution characteristics, improves robust- ness against cross-institutional domain shifts and provides better generalization than purely data-driven segmentation models. Table 4 Comparison between a registration pipeline and the proposed direct ADC-based bone ROI generation method. Best results are shown in bold. MetricRegistration Method Our Generation Method Speed per case∼43.5 s∼4.2 s Bone–Lesion Dice23.04 ± 13.0762.34 ± 26.96 Lesion recall22.6%70.9% 4.4. Ablation Studies 4.4.1. Effect of the Main AMU-Net Components Table 3 summarizes the ablation results of the main Stage-I design components in AMU-Net. In the structural ablation, removing either the encoder-level ROI-gated cross- attention or the decoder-level asymmetric gate leads to degraded performance compared with the full model, and removing both modules gives the worst Dice and IoU. This confirms that both components contribute positively to M lesion segmentation and provide complementary benefits. The larger performance drop caused by removing the encoder module suggests that encoder-level ADC inte- gration plays a particularly important role in supplement- ing structural information missing from WB-DWI, while the decoder-level asymmetric gate further improves lesion confirmation through DWI–ADC concordance modeling. 4.4.2. Effect of ROI Guidance The lower part of Table 3 evaluates the role of ROI guidance within the full model. Removing encoder-side ROI gating or decoder-side ROI amplification both reduces Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 9 of 11 performance, indicating that explicit skeletal prior benefits both structural supplementation and lesion confirmation. This effect is not only attributable to spatial restriction. As shown in Fig. 5, the average ADC value cannot serve as a reliable indicator of disease across patients, because marrow appearance can vary substantially with individual factors such as age and baseline marrow composition. As a result, lesion assessment is often better understood relative to the patient’s own skeletal marrow background rather than by absolute signal level alone. Bone ROI guidance therefore helps the model interpret DWI and ADC patterns within the patient’s marrow space, making it easier to distinguish focal lesion-related abnormalities from globally increased or individually variable marrow signal. 4.5. Practical Value of Direct Bone ROI Generation Beyond the final segmentation results, we further exam- ined the practical value of Stage-I bone ROI generation. In routine M imaging, Dixon is often acquired together with WB-DWI, and bone segmentation on Dixon is known to be highly accurate. We therefore used a Dixon-to-DWI regis- tration pipeline as a strong reference to compare ROI acqui- sition efficiency and lesion coverage. Despite this favorable reference, the proposed direct ADC-based ROI generation was both substantially faster and more lesion-relevant. It reduced the per-case ROI acquisition time from approxi- mately 43.5 s to 4.2 s, achieving a nearly 10× speedup, while improving bone–lesion overlap by 39.3 percentage points (Dice: 62.34 vs. 23.04) and lesion recall within the ROI by 48.3 percentage points (70.9% vs. 22.6%). It also showed more reliable case-level behavior, with 79.1% of cases achieving Dice > 0.5, compared with 0% for the registration-based reference, and no complete failures versus 3.0%. These results indicate that our ADC-based bone ROI generation method is faster, more accurate, and more prac- tical than the registration-based alternative for whole-body DWI analysis. 5. Conclusion In this work, we presented a two-stage framework for multiple myeloma lesion segmentation on whole-body DWI. As M is a relatively rare hematologic malignancy, well- annotated WB-DWI data are particularly difficult to acquire, making data-efficient model design especially important. At the same time, although WB-DWI is well suited for visualizing M lesions, automated segmentation on this modality remains highly challenging because of limited anatomical delineation and the low specificity of marrow hyperintensity. To address these challenges, the proposed framework maximizes the use of available multimodal infor- mation by generating bone ROI guidance directly from ADC and performing lesion segmentation with AMU-Net, which leverages clinically motivated DWI–ADC interaction for structural supplementation and lesion confirmation. Despite limited training data, the method achieves clear and consis- tent performance gains, supporting the value of combining bone-aware guidance with structured multimodal integration for M lesion segmentation. 6. Acknowlegments This work was supported in part by the National High Level Hospital Clinical Research Funding under Grant 2025- PUMCH-D-002, in part by the National Natural Science Foundation of China under Grant 82372051, in part by the Peking Union Medical College Hospital Talent Cultivation Program (Category B) under Grant UGG06278, in part by the Beijing Natural Science Foundation under Grant L222099, and in part by Damo Academy through Damo Academy Research Intern Program. CRediT authorship contribution statement Mengmeng Zhang: Conceptualization of this study, Methodology, Software, Writing – original draft. Shengqian Huang: Data curation, Validation. Junde Zhou: Data cu- ration. Xiaoping Wu: Validation. Hao Luo: Writing – re- view, editing. Jing Wang: Data curation. Yicheng Sun: Data curation. Jiao Li: Data curation. Haibo Zhang: Data curation. Sheng Xie: Data curation. Fan Wang: Resources. Qin Wang: Resources, Supervision, Funding acquisition. Huadan Xue: Resources, Supervision, Funding acquisition. Yisheng Lv: Resources, Supervision. Fei-yue Wang: Su- pervision. References [1] Almeida, S.D., Santinha, J., Oliveira, F.P., Ip, J., Lisitskaya, M., Lourenço, J., Uysal, A., Matos, C., João, C., Papanikolaou, N., 2020. Quantification of tumor burden in multiple myeloma by atlas-based semi-automatic segmentation of wb-dwi. Cancer Imaging 20, 6. [2] Baliyan, V., Das, C.J., Sharma, R., Gupta, A.K., 2016. Diffusion weighted imaging: technique and applications. World journal of radiology 8, 785. [3] Bauer, F., Kächele, J., Bernhard, J., Hajiyianni, M., Weinhold, N., Sauer, S., Grözinger, M., Raab, M.S., Mai, E.K., Weber, T.F., et al., 2025. Advanced automated model for robust bone marrow segmen- tation in whole-body mri. Academic radiology 32, 2824–2835. [4] Bonaffini, P.A., Ippolito, D., Casiraghi, A., Besostri, V., Franzesi, C.T., Sironi, S., 2015. Apparent diffusion coefficient maps integrated in whole-body mri examination for the evaluation of tumor response to chemotherapy in patients with multiple myeloma. Academic radiology 22, 1163–1171. [5] Candito, A., Dragan, A., Holbrey, R., Ribeiro, A., Donners, R., Mes- siou, C., Tunariu, N., Koh, D.M., Blackledge, M.D., 2025. A weakly- supervised deep learning model for fast localisation and delineation of the skeleton, internal organs, and spinal canal on whole-body diffusion-weighted mri (wb-dwi). Computer Methods and Programs in Biomedicine , 109043. [6] Candito, A., Holbrey, R., Ribeiro, A., Messiou, C., Tunariu, N., Koh, D.M., Blackledge, M.D., 2024. Deep learning for delineation of the spinal canal in whole-body diffusion-weighted imaging: normalising inter-and intra-patient intensity signal in multi-centre datasets. Bio- engineering 11, 130. [7] Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al., 2022. Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 . Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 10 of 11 [8] Chen, Q., Zhang, J., Meng, R., Zhou, L., Li, Z., Feng, Q., Shen, D., 2024. Modality-specific information disentanglement from multi- parametric mri for breast tumor segmentation and computer-aided diagnosis. IEEE Transactions on Medical Imaging 43, 1958–1971. [9] Fan, T., Wang, G., Li, Y., Wang, H., 2020. Ma-net: A multi-scale attention network for liver and tumor segmentation. Ieee Access 8, 179656–179665. [10] Giles, S., Desouza, N., Collins, D., Morgan, V.A., West, S., Davies, F., Morgan, G., Messiou, C., 2015. Assessing myeloma bone disease with whole-body diffusion-weighted imaging: comparison with x-ray skeletal survey by region and relationship with laboratory estimates of disease burden. Clinical radiology 70, 614–621. [11] Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D., 2021. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images, in: International MICCAI brainlesion workshop, Springer. p. 272–284. [12] Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D., 2022. Unetr: Transformers for 3d medical image segmentation, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, p. 574–584. [13] He, Y., Guo, P., Tang, Y., Myronenko, A., Nath, V., Xu, Z., Yang, D., Zhao, C., Simon, B., Belue, M., et al., 2025. Vista3d: A unified segmentation foundation model for 3d medical imaging, in: Proceed- ings of the Computer Vision and Pattern Recognition Conference, p. 20863–20873. [14] Horger, M., Weisel, K., Horger, W., Mroue, A., Fenchel, M., Lichy, M., 2011. Whole-body diffusion-weighted mri with apparent diffu- sion coefficient mapping for early response monitoring in multiple myeloma: preliminary results. American Journal of Roentgenology 196, W790–W795. [15] Ibtehaz, N., Rahman, M.S., 2020. Multiresunet: Rethinking the u-net architecture for multimodal biomedical image segmentation. Neural networks 121, 74–87. [16] Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H., 2021. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18, 203–211. [17] Koutoulakis, E., Trivizakis, E., Koutoulidis, V., Moulopoulos, L.A., Terpos, E., Ntanasis-Stathopoulos, I., Malandrakis, P., Grigoropou- los, P., Papadopoulos, P., Nikiforaki, K., et al., 2025. Label-free machine learning-based segmentation of whole-body bone marrow imaging in multiple myeloma, in: 2025 IEEE 19th International Symposium on Applied Computational Intelligence and Informatics (SACI), IEEE. p. 1–6. [18] Messiou, C., Kaiser, M., 2015. Whole body diffusion weighted mri–a new view of myeloma. British journal of haematology 171, 29–37. [19] Michoux, N.F., Ceranka, J.W., Vandemeulebroucke, J., Peeters, F., Lu, P., Absil, J., Triqueneaux, P., Liu, Y., Collette, L., Willekens, I., et al., 2021. Repeatability and reproducibility of adc measurements: a prospective multicenter whole-body-mri study. European radiology 31, 4514–4527. [20] Mo, S., Cai, M., Lin, L., Tong, R., Chen, Q., Wang, F., Hu, H., Iwamoto, Y., Han, X.H., Chen, Y.W., 2021. Mutual information-based graph co-attention networks for multimodal prior-guided magnetic resonance imaging segmentation. IEEE Transactions on Circuits and Systems for Video Technology 32, 2512–2526. [21] Myronenko, A., 2018. 3d mri brain tumor segmentation using autoen- coder regularization, in: International MICCAI brainlesion workshop, Springer. p. 311–320. [22] Nguyen, C.T., Havlicek, J.P., Chakrabarty, J.H., Duong, Q., Vesely, S.K., 2016. Towards automatic 3d bone marrow segmentation, in: 2016 IEEE Southwest Symposium on Image Analysis and Interpreta- tion (SSIAI), IEEE. p. 9–12. [23] Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Mis- awa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al., 2018. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 . [24] Paternain, A., García-Velloso, M.J., Rosales, J.J., Ezponda, A., Sori- ano, I., Elorz, M., Rodríguez-Otero, P., Aquerreta, J.D., 2020. The utility of adc value in diffusion-weighted whole-body mri in the follow-up of patients with multiple myeloma. correlation study with 18f-fdg pet-ct. European Journal of Radiology 133, 109403. [25] Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation, in: International Con- ference on Medical image computing and computer-assisted interven- tion, Springer. p. 234–241. [26] Summers, P., Saia, G., Colombo, A., Pricolo, P., Zugni, F., Alessi, S., Marvaso, G., Jereczek-Fossa, B.A., Bellomi, M., Petralia, G., 2021. Whole-body magnetic resonance imaging: technique, guidelines and key applications. Ecancermedicalscience 15, 1164. [27] Tiribilli, E., Iadanza, E., Bocchi, L., 2024. Single bone modeler: Deep learning bone segmentation for cone-beam ct, in: 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), IEEE. p. 1–4. [28] Torkian, P., Mansoori, B., Hillengass, J., Azadbakht, J., Rashedi, S., Lee, S.S., Amini, B., Bonaffini, P.A., Chalian, M., 2023. Diffusion- weighted imaging (dwi) in diagnosis, staging, and treatment response assessment of multiple myeloma: a systematic review and meta- analysis. Skeletal radiology 52, 565–583. [29] Valanarasu, J.M.J., Patel, V.M., 2022. Unext: Mlp-based rapid medical image segmentation network, in: International conference on medical image computing and computer-assisted intervention, Springer. p. 23–33. [30] Varghese, R., Sharma, S., Premalatha, M., 2018. Transforming auto- encoder and decoder network for pediatric bone image segmentation using a state-of-the-art semantic segmentation network on bone radio- graphs, in: 2018 International Conference on Intelligent Informatics and Biomedical Sciences (ICIIBMS), IEEE. p. 251–256. [31] Wennmann, M., Neher, P., Stanczyk, N., Kahl, K.C., Kächele, J., Weru, V., Hielscher, T., Grözinger, M., Chmelik, J., Zhang, K.S., et al., 2023. Deep learning for automatic bone marrow apparent diffusion coefficient measurements from whole-body magnetic res- onance imaging in patients with multiple myeloma: a retrospective multicenter study. Investigative radiology 58, 273–282. [32] Wu, C., Huang, J., Xu, W.B., Guan, Y.J., Ling, H.W., Mi, J.Q., Yan, H., 2018. Discriminating depth of response to therapy in multiple myeloma using whole-body diffusion-weighted mri with apparent diffusion coefficient: preliminary results from a single-center study. Academic radiology 25, 904–914. [33] Wuts, J., Ceranka, J., Michoux, N., Lecouvet, F., Vandemeulebroucke, J., 2026. Clinically aligned whole-body mri segmentation of skeletal metastases via supervised anatomical pretraining. Journal of Bone Oncology , 100745. [34] Zhao, X., Chen, L., Zhang, N., Lv, Y., Hu, X., 2024. Multiple myeloma segmentation net (mmnet): an encoder-decoder-based deep multiscale feature fusion model for multiple myeloma segmentation in magnetic resonance imaging. Quantitative Imaging in Medicine and Surgery 14, 7176–7199. [35] Zhuang, Y., Liu, H., Fang, W., Ma, G., Sun, S., Zhu, Y., Zhang, X., Ge, C., Chen, W., Long, J., et al., 2024. A 3d hierarchical cross- modality interaction network using transformers and convolutions for brain glioma segmentation in mr images. Medical Physics 51, 8371– 8389. [36] Zhuang, Y., Liu, H., Song, E., Hung, C.C., 2022. A 3d cross-modality feature interaction network with volumetric feature alignment for brain tumor and tissue segmentation. IEEE Journal of Biomedical and Health Informatics 27, 75–86. Mengmeng Zhang et.al : Preprint submitted to ElsevierPage 11 of 11