Paper deep dive
STRIDE: Spatial-Temporal Representation for Interval-conditioned Disease Evolution in Longitudinal Glioblastoma MRI
Wenhao Guo, Changchang Yin, Pierre Giglio, Weidan Cao, Ping Zhang, Golrokh Mirzaei
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 10/8/2026, 6:05:50 AM
Summary
The paper introduces STRIDE, a spatial-temporal representation framework for interval-conditioned disease evolution in longitudinal Glioblastoma (GBM) MRI. STRIDE predicts Stable Disease (SD), Pseudoprogression (PsP), and True Progression (TP) by integrating lesion-prior-guided spatial representation, time-conditioned latent transition, and observed-transition fusion. It addresses challenges in scan-specific spatial variability, variable follow-up intervals, and complementary information from observed states and longitudinal changes. Evaluated on the Burdenko dataset, STRIDE achieves a macro ROC-AUC of 0.816 and macro F1-score of 0.796.
Entities (11)
Relation Signals (9)
STRIDE → evaluatedon → Burdenko
confidence 95% · On the Burdenko three-class task, STRIDE achieves a macro ROC--AUC of 0.816
Glioblastoma → monitoredby → MRI
confidence 95% · Glioblastoma (GBM)... is routinely monitored with longitudinal MRI after treatment.
STRIDE → predicts → Pseudoprogression
confidence 95% · predicts SD, PsP, or TP
STRIDE → predicts → True Progression
confidence 95% · predicts SD, PsP, or TP
STRIDE → predicts → Stable Disease
confidence 95% · predicts SD, PsP, or TP
STRIDE → usescomponent → SoftGate
confidence 95% · The lesion-prior-guided spatial representation combines SoftGate and an Adaptive-window Hierarchical Transformer
STRIDE → usescomponent → Adaptive-window Hierarchical Transformer
confidence 95% · The lesion-prior-guided spatial representation combines SoftGate and an Adaptive-window Hierarchical Transformer (AWHT)
BraTS2024 → usedfor → lesion-prior generator
confidence 90% · BraTS2024 is used to develop and evaluate the lesion-prior generator
→ →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Glioblastoma (GBM), an aggressive primary brain tumor, is routinely monitored with longitudinal MRI after treatment. Distinguishing stable disease (SD), pseudoprogression (PsP), and true progression (TP) remains challenging because these states can show overlapping MRI appearances despite different temporal trajectories. Existing longitudinal methods still face challenges in modeling scan-specific spatial variability, variable follow-up intervals, and complementary information from the observed follow-up state and its longitudinal change. We propose STRIDE, a framework for spatial-temporal representation of interval-conditioned disease evolution that takes paired post-treatment MRI scans and their inter-scan interval as input and predicts SD, PsP, or TP. The lesion-prior-guided spatial representation combines SoftGate and an Adaptive-window Hierarchical Transformer (AWHT) to emphasize lesion-related regions while preserving surrounding context. The time-conditioned latent transition uses pair-level context and the actual inter-scan interval to estimate interval-dependent representation changes between visits. The observed--transition fusion integrates the transition-estimated follow-up representation with the directly observed follow-up representation to jointly characterize the follow-up state and its longitudinal change. BraTS2024 is used to develop and evaluate the lesion-prior generator, while longitudinal pretraining on LUMIERE supports transfer before downstream adaptation to Burdenko. On the Burdenko three-class task, STRIDE achieves a macro ROC--AUC of 0.816 and a macro F1-score of 0.796. These results support its potential for more reliable longitudinal post-treatment GBM state assessment.
Tags
Links
- Source: https://arxiv.org/abs/2610.08848v1
- Canonical: https://arxiv.org/abs/2610.08848v1
Trouble viewing inline? Open PDF directly →
Full Text
73,810 characters extracted from source content.
Expand or collapse full text
STRIDE: Spatial-Temporal Representation for Interval-conditioned Disease Evolution in Longitudinal Glioblastoma MRI Journal: Medical Image Analysis Wenhao Guo Note: These authors contributed equally to this work. Affiliation: Department of Computer Science and Engineering, The Ohio State University, Columbus, OH, USA Changchang Yin Note: These authors contributed equally to this work. Affiliation: Department of Biomedical Informatics, The Ohio State University, Columbus, OH, USA Pierre Giglio Affiliation: Department of Neurology, The Ohio State University, Columbus, OH, USA Weidan Cao Affiliation: Department of Biomedical Informatics, The Ohio State University, Columbus, OH, USA Ping Zhang Email: zhang.10631@osu.edu Corresponding author: Corresponding author. Affiliation: Department of Computer Science and Engineering, The Ohio State University, Columbus, OH, USA Affiliation: Department of Biomedical Informatics, The Ohio State University, Columbus, OH, USA Golrokh Mirzaei Email: mirzaei.4@osu.edu Corresponding author: Corresponding author. Affiliation: Department of Computer Science and Engineering, The Ohio State University, Columbus, OH, USA Abstract Glioblastoma (GBM), an aggressive primary brain tumor, is routinely monitored with longitudinal MRI after treatment. Distinguishing stable disease (SD), pseudoprogression (PsP), and true progression (TP) remains challenging because these states can show overlapping MRI appearances despite different temporal trajectories. Existing longitudinal methods still face challenges in modeling scan-specific spatial variability, variable follow-up intervals, and complementary information from the observed follow-up state and its longitudinal change. We propose STRIDE, a framework for spatial-temporal representation of interval-conditioned disease evolution that takes paired post-treatment MRI scans and their inter-scan interval as input and predicts SD, PsP, or TP. The lesion-prior-guided spatial representation combines SoftGate and an Adaptive-window Hierarchical Transformer (AWHT) to emphasize lesion-related regions while preserving surrounding context. The time-conditioned latent transition uses pair-level context and the actual inter-scan interval to estimate interval-dependent representation changes between visits. The observed–transition fusion integrates the transition-estimated follow-up representation with the directly observed follow-up representation to jointly characterize the follow-up state and its longitudinal change. BraTS2024 is used to develop and evaluate the lesion-prior generator, while longitudinal pretraining on LUMIERE supports transfer before downstream adaptation to Burdenko. On the Burdenko three-class task, STRIDE achieves a macro ROC–AUC of 0.816 and a macro F1-score of 0.796. These results support its potential for more reliable longitudinal post-treatment GBM state assessment. Keywords: Glioblastoma , Longitudinal MRI , Pseudoprogression , Temporal Modeling , Lesion-aware Learning 1 Introduction Glioblastoma (GBM) is the most aggressive primary brain tumor in adults and remains a major challenge in neuro-oncology. Despite advances in surgery, radiotherapy, and chemotherapy, tumor recurrence and progression after treatment remain common, leading to poor clinical outcomes (Stupp et al., 2005). Longitudinal magnetic resonance imaging (MRI) is routinely used to monitor treatment response and guide clinical decisions, with standardized criteria such as RANO and its updated versions supporting imaging-based response assessment (Wen et al., 2010; Okada et al., 2015; Okada et al., 2018; Wen et al., 2023). However, reliable differentiation among stable disease (SD), pseudoprogression (PsP), and true progression (TP) remains challenging because post-treatment imaging appearances can overlap while the corresponding disease trajectories differ (Wen et al., 2010; Taal et al., 2008; Brandes et al., 2008). Distinguishing PsP from TP is especially difficult during early post-treatment follow-up, making longitudinal assessment important for reliable clinical interpretation (Taal et al., 2008; Brandes et al., 2008; Young et al., 2023). Recent artificial intelligence approaches have demonstrated potential for automated GBM progression assessment, including radiomics and classical machine learning (Turk et al., 2022; Jang et al., 2018; Kim et al., 2019; Elshafeey et al., 2019; Akbari et al., 2020), volumetric and multimodal deep learning (Moassefi et al., 2022; Li et al., 2025), Transformer-based architectures (Guo and Mirzaei, 2026), and self-supervised representation learning (Gomaa et al., 2025). Despite this progress, systematic reviews and meta-analyses have reported substantial variability in diagnostic performance across cohorts (Palavani et al., 2025; Kotochinsky et al., 2026). One persistent challenge is that a single MRI examination provides only a snapshot of the disease state and does not directly capture how the disease changes between visits. This has motivated recurrent models (Siakallis et al., 2021; Cui et al., 2019), Siamese and difference-based learning (Li et al., 2020; Zheng et al., 2025), and temporal attention and Transformer-based longitudinal approaches (Aghajanian et al., 2025; Wang et al., 2024; Nguyen et al., 2023; Sun et al., 2024; Zhang et al., 2025; Zheng et al., 2026; Tarek et al., 2026). However, incorporating multiple visits does not by itself resolve three key modeling challenges. First, progression-related abnormalities vary in location and spatial extent across MRI scans, while existing longitudinal representations remain limited in adapting to such scan-specific spatial variability (Guo and Mirzaei, 2026; Gomaa et al., 2025). Second, follow-up intervals vary substantially across patients, yet the actual inter-scan interval remains insufficiently incorporated in existing longitudinal modeling of post-treatment GBM (Chen et al., 2018; Rubanova et al., 2019; Che et al., 2018; Ghazi et al., 2019; Mouadden et al., 2025; De Brouwer et al., 2019; Liu et al., 2025; Tarek et al., 2026). Third, the observed follow-up state and the longitudinal change leading to that state provide complementary information, while existing longitudinal models remain limited in jointly characterizing these two aspects of disease evolution (Cui et al., 2019; Li et al., 2020; Sun et al., 2024; Zhang et al., 2025; Zheng et al., 2026; Tarek et al., 2026). To address these challenges, we propose STRIDE, a spatial-temporal representation framework for interval-conditioned disease evolution that takes paired post-treatment MRI scans and their inter-scan interval as input and predicts SD, PsP, or TP. For each MRI scan, the lesion-prior-guided spatial representation uses segmentation-derived structural guidance (Ronneberger et al., 2015; Isensee et al., 2021; Hatamizadeh et al., 2022b; Hatamizadeh et al., 2022a; Mehta et al., 2018; Fang et al., 2019; Gao et al., 2026) through a mask-conditioned SoftGate to emphasize lesion-supported features while retaining surrounding context. An Adaptive-window Hierarchical Transformer (AWHT) then combines scan-adaptive multi-scale localization with hierarchical encoding to produce a visit-level spatial representation. The spatial representations from the two visits are subsequently used to construct pair-level longitudinal context. The time-conditioned latent transition combines the earlier-visit representation, pair-level context, and the actual inter-scan interval to estimate interval-dependent representation changes and produce a transition-estimated follow-up representation. The observed–transition fusion then integrates the earlier-visit representation, the directly observed and transition-estimated follow-up representations, and their associated longitudinal information to jointly characterize the follow-up state and its longitudinal change. Because annotated longitudinal GBM cohorts remain limited, longitudinal pretraining is additionally employed to improve representation transfer before downstream adaptation (Zhou et al., 2021; Tang et al., 2022; Huo et al., 2025; Huang et al., 2026; Gomaa et al., 2025; Suter et al., 2022). Our contributions are summarized as follows: 1. We propose STRIDE, a longitudinal Transformer framework for three-class post-treatment GBM clinical-state classification that jointly addresses scan-specific spatial heterogeneity, variable inter-scan intervals, and complementary state–change information. 2. We introduce lesion-prior-guided spatial representation that combines structural guidance with scan-adaptive multi-scale localization to emphasize lesion-related regions while preserving surrounding context. 3. We develop a time-conditioned latent transition that incorporates pair-level context and the actual inter-scan interval to explicitly model interval-dependent representation changes between visits. 4. We design an observed–transition fusion that integrates directly observed and transition-estimated follow-up representations to capture complementary state–change information for SD, PsP, and TP classification. 2 Proposed Method 2.1 Overview of STRIDE Framework STRIDE consists of three major components: lesion-prior-guided spatial representation, time-conditioned latent transition, and observed–transition fusion. These components progressively extract visit-specific spatial representations, characterize interval-dependent changes between scans, and integrate observed and transition-estimated longitudinal information for three-class clinical-state classification. The overall architecture is illustrated in Figure 1. Figure 1: Overall architecture of STRIDE. Given paired longitudinal MRI scans and their inter-scan interval, segmentation-derived lesion priors first provide structural guidance for lesion-prior-guided spatial representation. The resulting visit-specific representations are used to construct pair-level context, while the inter-scan interval is encoded as temporal information. These representations are then processed by the time-conditioned latent transition to characterize interval-dependent latent changes and obtain a transition-estimated follow-up representation. Finally, the observed and transition-estimated longitudinal representations are jointly fused for three-class classification of stable disease (SD), pseudoprogression (PsP), and true progression (TP). 2.2 Problem Formulation For a patient with N longitudinal MRI scans, we denote the imaging sequence as =(Xi,ti)i=1NX=\(X_i,t_i)\_i=1^N, where Xi∈ℝM×H×W×DX_i ^M× H× W× D is the multimodal MRI scan acquired at visit i, tit_i is its acquisition time, M denotes the number of imaging modalities, and H, W, and D denote the spatial dimensions. Although a patient may have multiple MRI scans, STRIDE operates on visit pairs. For visits i<ji<j, the corresponding inter-scan interval is Δtij=tj−ti t_ij=t_j-t_i, and each sample is represented by (Xi,Xj,Δtij)(X_i,X_j, t_ij). For downstream classification, each pair is assigned the patient’s curated clinical-state label cij∈SD,PsP,TPc_ij∈\SD,PsP,TP\ and its corresponding one-hot vector ij∈0,13y_ij∈\0,1\^3. Given a longitudinal pair, STRIDE predicts ij=Fθ(Xi,Xj,Δtij),∑k=13pij,k=1,p_ij=F_θ (X_i,X_j, t_ij ), _k=1^3p_ij,k=1, (1) where ij∈[0,1]3p_ij∈[0,1]^3, and its three elements correspond to SD, PsP, and TP, respectively. The predicted clinical state is c^ij=argmaxkpij,k c_ij= _kp_ij,k. 2.3 Lesion-prior-guided Spatial Representation Post-treatment abnormalities can be concentrated within localized lesion regions, while surrounding tissue can still contain information relevant to distinguishing SD, PsP, and TP. Treating all spatial locations equally makes lesion-related signals less prominent, whereas restricting representation learning to a hard lesion mask removes useful contextual features. We therefore combine segmentation-derived structural guidance with soft lesion-aware modulation and adaptive spatial representation learning. For each visit, a frozen segmentation network provides a lesion prior. SoftGate uses this prior to modulate image features, and the resulting representation is processed by the Adaptive-window Hierarchical Transformer (AWHT) with scan-adaptive spatial emphasis. The module is illustrated in Figure 2. Figure 2: Lesion-prior-guided spatial representation. For each visit, a segmentation-derived lesion prior provides structural guidance to the mask-conditioned SoftGate, which emphasizes lesion-supported features while retaining surrounding contextual information. The resulting representation is processed by the Adaptive-window Hierarchical Transformer (AWHT). AWHT constructs an input-dependent continuous spatial window from multi-scale localization responses and subsequently encodes the weighted representation through hierarchical window-based self-attention, producing the visit-level representation vz_v. The spatial representation pathway is shared across longitudinal visits. 2.3.1 Segmentation-derived Lesion Prior For visit v, a pretrained segmentation network generates a voxel-wise abnormal-region probability map Mv=S(Xv)M_v=S(X_v), where Mv∈[0,1]1×H×W×DM_v∈[0,1]^1× H× W× D. The segmentation network remains frozen during longitudinal learning, and its output is used only as auxiliary structural guidance for subsequent spatial representation learning. 2.3.2 Mask-conditioned SoftGate SoftGate applies the lesion prior as a soft emphasis signal while preserving the original feature representation. Let v∈ℝHF×WF×DF×CFF_v ^H_F× W_F× D_F× C_F denote the visit-specific feature map used by SoftGate. After spatial alignment, the lesion prior is projected by ϕ(⋅)φ(·) to a representation compatible with vF_v. A continuous gating map is then obtained as v=σ(v⊙ϕ(Mv)),G_v=σ (F_v φ(M_v) ), (2) where ϕ(⋅)φ(·) denotes a learnable prior projection, σ(⋅)σ(·) denotes the sigmoid activation, and ⊙ denotes element-wise multiplication. The gate is applied through residual modulation: ~v=v+αv⊙v, F_v=F_v+α\,F_v _v, (3) where α∈[0,1]α∈[0,1] controls the strength of lesion-aware modulation. In all experiments, α was fixed to 0.5. 2.3.3 Adaptive-window Hierarchical Transformer To accommodate variations in lesion location and spatial extent, AWHT constructs an input-dependent continuous spatial window from multi-scale image features before hierarchical window-based Transformer encoding. It contains an input-adaptation stem, multi-scale adaptive spatial-window estimation, and hierarchical 3D window-based representation encoding. Here, adaptive window refers to the examination-specific continuous spatial weighting field, while the local self-attention window remains fixed at 8×8×88× 8× 8 tokens. Input adaptation The SoftGate-modulated representation is first projected into a common feature space using two consecutive 3×3×33× 3× 3 convolutional blocks: v _v =Dropout3D[ψ2(ψ1(~v))], =Dropout3D [ _2 ( _1( F_v) ) ], (4) ψr() _r(F) =GELU[BNr(Convr3×3×3())]. =GELU [BN_r (Conv^3× 3× 3_r(F) ) ]. where v∈ℝHF×WF×DF×C0U_v ^H_F× W_F× D_F× C_0 and C0C_0 is the stem feature dimension. Multi-scale adaptive spatial window Because abnormalities can occupy substantially different spatial extents, localization cues obtained from a single receptive-field scale are insufficient. We therefore use three parallel localization branches operating at scales =1,2,4S=\1,2,4\: v(s)=Convs,21×1×1[ReLU(BNs[Convs,13×3×3(v)])],L^(s)_v=Conv^1× 1× 1_s,2 [ReLU (BN_s [Conv^3× 3× 3_s,1(U_v) ] ) ], (5) where the first convolution uses stride s∈s . Larger strides provide localization cues over broader effective receptive fields. The scale-specific responses are subsequently aligned to the spatial resolution of vU_v and aggregated to obtain the adaptive spatial window v=σ[1||∑s∈ℐs(v(s))],A_v=σ [ 1|S| _s I_s (L^(s)_v ) ], (6) where v∈[0,1]HF×WF×DF×1A_v∈[0,1]^H_F× W_F× D_F× 1 and ℐs(⋅)I_s(·) denotes trilinear interpolation to the common spatial resolution. The adaptive spatial window is then applied through element-wise weighting, ~v=v⊙v U_v=U_v _v, where vA_v is broadcast across feature channels and ~v∈ℝHF×WF×DF×C0 U_v ^H_F× W_F× D_F× C_0. This produces an examination-specific weighted feature volume for hierarchical representation encoding. Hierarchical window-based representation encoding The weighted feature volume ~v U_v is divided into non-overlapping 4×4×44× 4× 4 patches and projected to an initial embedding dimension C0=96C_0=96: v(0)=PatchEmbed(~v),Z^(0)_v=PatchEmbed ( U_v ), (7) where v(0)∈ℝN0×C0Z^(0)_v ^N_0× C_0 and N0N_0 denotes the number of initial 3D patch tokens. The hierarchical encoder contains four stages with block depths (2,2,2,2)(2,2,2,2) and attention-head configurations (3,6,12,24)(3,6,12,24). Let v(l,b−1)∈ℝNl×ClZ^(l,b-1)_v ^N_l× C_l denote the input representation to block b of stage l, where NlN_l is the number of tokens and ClC_l is the corresponding embedding dimension. Within each stage, regular window-based self-attention and shifted-window self-attention are alternated: ^v(l,b) Z^(l,b)_v =v(l,b−1)+(l,b)[LN(v(l,b−1))], =Z^(l,b-1)_v+A^(l,b) [LN (Z^(l,b-1)_v ) ], (8) v(l,b) ^(l,b)_v =^v(l,b)+MLP[LN(^v(l,b))], = Z^(l,b)_v+MLP [LN ( Z^(l,b)_v ) ], (9) where ^v(l,b),v(l,b)∈ℝNl×Cl Z^(l,b)_v,Z^(l,b)_v ^N_l× C_l, and (l,b)A^(l,b) alternates between regular and shifted-window self-attention. Patch merging between consecutive stages reduces the number of spatial tokens while increasing the embedding dimension, producing v(l+1,0)∈ℝNl+1×Cl+1Z^(l+1,0)_v ^N_l+1× C_l+1. After the final hierarchical stage, global average pooling and normalization produce the visit-level representation v=LN[GAP(v(L−1))].z_v=LN [GAP (Z^(L-1)_v ) ]. (10) where v∈ℝZz_v ^Z. Applying the same spatial encoder to both examinations yields iz_i and j∈ℝZz_j ^Z for subsequent longitudinal modeling. 2.4 Time-conditioned Latent Transition The spatial module provides representations of the two observed MRI scans. To characterize their relationship over irregular follow-up intervals, the time-conditioned latent transition combines inter-scan interval encoding, pair-level context modeling, and gated transition estimation (Figure 3). Figure 3: Time-conditioned latent transition. The inter-scan interval Δtij t_ij is first log-scaled, periodically encoded, and mapped to the interval embedding ije_ij. The observed visit representations iz_i and jz_j are then used to compute their latent difference and adaptively weighted cross-visit interaction, which together form the pair-level context pairc_pair. The earlier representation, pair-level context, and interval embedding are subsequently used to estimate an interval-scaled gated latent change, producing the transition-estimated representation ^j z_j. 2.4.1 Inter-scan Interval Encoding To condition transition modeling on follow-up timing, the inter-scan interval is log-scaled and mapped to a temporal embedding: Δt¯ij=log(1+Δtijτ),ij=ftime[γ(Δt¯ij)], t_ij= (1+ t_ijτ ), _ij=f_time [γ( t_ij) ], (11) where ij∈ℝEe_ij ^E, τ>0τ>0 is the temporal scaling constant, γ(⋅)γ(·) combines the log-scaled interval with periodic temporal features, and ftime(⋅)f_time(·) is a multilayer perceptron. The resulting ije_ij conditions the subsequent latent transition. 2.4.2 Pair-level Context Modeling Pair-level context summarizes the observed visit representations, their change, and their cross-visit feature interaction. The observed latent change is defined as Δobs=j−i. ^obs=z_j-z_i. (12) where Δobs∈ℝZ ^obs ^Z. This difference captures additive change but not interactions between corresponding latent dimensions. We therefore compute an adaptively weighted cross-visit interaction: =σ(α[i;j;Δobs]+α),int=⊙(i⊙j). α=σ (W_α [z_i;z_j; ^obs ]+b_α ), _int= α (z_i _j ). (13) where ∈(0,1)Z α∈(0,1)^Z, int∈ℝZh_int ^Z, α∈ℝZ×3ZW_α ^Z× 3Z, and α∈ℝZb_α ^Z are learnable parameters. The pair-level context then combines the observed visit representations with the observed latent change and weighted cross-visit interaction: pair=fctx[i;j;Δobs;int].c_pair=f_ctx [z_i;z_j; ^obs;h_int ]. (14) where pair∈ℝCc_pair ^C and fctx(⋅)f_ctx(·) is a multilayer perceptron. 2.4.3 Gated Latent Transition The earlier representation, pair-level context, and temporal embedding are combined to estimate a gated transition over the inter-scan interval. A transition representation dyn=fdyn[i;pair;ij]h_dyn=f_dyn[z_i;c_pair;e_ij] is first computed, where dyn∈ℝHh_dyn ^H and fdyn(⋅)f_dyn(·) denotes the transition mapping. The candidate latent change and its element-wise gate are then computed as Δ=Δdyn,=σ(gdyn), =W_ h_dyn, =σ (W_gh_dyn ), (15) where Δ∈ℝZ ^Z, ∈(0,1)Zg∈(0,1)^Z, and Δ,g∈ℝZ×HW_ ,W_g ^Z× H. The transition-estimated follow-up representation is then computed as ^j=i+Δt¯ij(⊙Δ). z_j=z_i+ t_ij (g ). (16) where ^j∈ℝZ z_j ^Z. Here, iz_i provides the transition origin, g modulates the contribution of each latent dimension to the update, and Δt¯ij t_ij scales the estimated change according to the inter-scan interval. The resulting ^j z_j provides a transition-estimated follow-up representation that complements the directly observed follow-up representation. Because the pair-level context is constructed from both observed visits, ^j z_j is interpreted as a pair-conditioned representation of interval-dependent change rather than a future-state prediction derived from iz_i alone. 2.5 Observed–Transition Fusion The directly observed representation j∈ℝZz_j ^Z and transition-estimated representation ^j∈ℝZ z_j ^Z provide complementary information about the follow-up state. The fusion module integrates them with cross-visit information for classification (Figure 4). Figure 4: Observed–transition fusion. The fusion module integrates the earlier representation iz_i, observed follow-up representation jz_j, transition-estimated representation ^j z_j, observed and estimated latent changes, pair-level context pairc_pair, and interval embedding ije_ij. These longitudinal sources are first mapped into a common representation space and integrated through multi-source cross-attention. The resulting information is aggregated into the fused representation fusionh_fusion, which is refined and used for three-class classification of SD, PsP, and TP. The transition-estimated change is defined as Δest=^j−i. ^est= z_j-z_i. (17) where Δest∈ℝZ ^est ^Z. To preserve both state and change information, the observed and transition-estimated visit representations, their corresponding latent changes, pair-level context, and temporal embedding are projected into a common representation space: 0=[1(i);2(j);3(^j);4(Δobs);5(Δest);6(pair);7(ij)],S_0= [P_1(z_i);P_2(z_j);P_3( z_j);P_4( ^obs);P_5( ^est);P_6(c_pair);P_7(e_ij) ], (18) where 0∈ℝNS×ZS_0 ^N_S× Z, k(⋅)P_k(·) denotes a learnable source-specific projection, and NSN_S denotes the number of projected source tokens. The projected sources are then processed by the learnable longitudinal fusion operator ℱfusionF_fusion and aggregated into a unified representation: fusion=Agg[ℱfusion(0)],h_fusion=Agg [F_fusion (S_0 ) ], (19) where fusion∈ℝFh_fusion ^F and Agg(⋅)Agg(·) denotes the fusion aggregation operation. Consistent with the prediction head in Figure 4, the fused representation is further refined by an MLP fref:ℝF→ℝRf_ref:R^F ^R, producing ref∈ℝRh_ref ^R. The refined representation is then passed to a learnable classification head Cls:ℝR→ℝKCls:R^R ^K, followed by softmax, to obtain the probability vector ij∈[0,1]Kp_ij∈[0,1]^K. 2.6 Training Objectives The prediction output is optimized using a class-weighted classification objective together with transition-specific regularization. Because LUMIERE pretraining and Burdenko downstream adaptation use different label spaces, the class-weighted cross-entropy loss is defined as ℒcls=−∑k=1Kwkyij,klogpij,k,L_cls=- _k=1^Kw_ky_ij,k p_ij,k, (20) where K denotes the number of classes at the corresponding training stage, with K=7K=7 for LUMIERE pretraining and K=3K=3 for Burdenko downstream adaptation. Here, yij,ky_ij,k is the target indicator for class k and wkw_k is the corresponding class weight. Because the transition-estimated representation is intended to remain consistent with the observed follow-up representation, we further constrain their latent-space discrepancy using ℒtr=‖^j−j‖22,ℒdir=1−cos(^j,j),L_tr= \| z_j-z_j \|_2^2, _dir=1- ( z_j,z_j ), (21) where cos(⋅,⋅) (·,·) denotes cosine similarity. The transition loss ℒtrL_tr constrains the magnitude of the discrepancy between the transition-estimated and observed follow-up representations, whereas the latent directional consistency loss ℒdirL_dir encourages directional alignment between the two follow-up representations in latent space. The complete objectives for longitudinal pretraining and downstream adaptation are defined as ℒpre _pre =ℒcls+λtrℒtr+λdirℒdir, =L_cls+ _trL_tr+ _dirL_dir, (22) ℒdown _down =ℒcls+βtrℒtr+βdirℒdir, =L_cls+ _trL_tr+ _dirL_dir, (23) where ℒpreL_pre and ℒdownL_down denote the objectives used for LUMIERE pretraining and Burdenko downstream adaptation, respectively. λtr _tr and λdir _dir control the contributions of the transition and latent directional consistency losses during pretraining, whereas βtr _tr and βdir _dir control their contributions during downstream adaptation. 3 Experimental Setup 3.1 Datasets and Data Preparation We used BraTS2024 (de Verdier et al., 2024) for lesion-prior generation, LUMIERE (Suter et al., 2022) for longitudinal pretraining, and the Burdenko Glioblastoma Progression Dataset (Zolotova et al., 2023) for downstream three-class clinical-state classification (Table 1). Table 1: Characteristics and study roles of the three datasets. Age (mean ± standard deviation) and sex are summarized per patient, whereas clinical-label counts are reported per longitudinal pair. Statistics for LUMIERE and Burdenko correspond to the retained cohorts after quality control and longitudinal sample construction. Abbreviations: M/F, male/female; WT, whole tumor; TC, tumor core; ET, enhancing tumor; RC, resection cavity; PD, progressive disease; SD, stable disease; PR, partial response; CR, complete response; PsP, pseudoprogression; TP, true progression; N/A, not applicable. Characteristic BraTS2024 LUMIERE Burdenko Study role Lesion-prior generation Longitudinal pretraining Downstream classification Patients 613 87 98 Age (years) 52.9 ± 14.1 62.1 ± 10.3 54.5 ± 12.2 Sex (M/F) 355/258 44/43 50/48 MRI scans 1,350 591 279 Mean scans/patient 2.20 6.79 2.85 Longitudinal pairs N/A 499 181 Clinical labels (pairs) N/A PD, 233 (46.7%); SD, 93 (18.6%); PR, 20 (4.0%); CR, 25 (5.0%); post-operative, 99 (19.8%); pre-operative, 1 (0.2%); unknown, 28 (5.6%) SD, 39 (21.5%); PsP, 31 (17.1%); TP, 111 (61.3%) Segmentation WT, TC, ET, and RC Automated tumor masks N/A Temporal information N/A Longitudinal scan timing Longitudinal follow-up timing We used all 1,350 annotated examinations from 613 patients in the initial BraTS2024 Adult Glioma Post-Treatment training release, each with MNI-space T1n, T1c, T2w, T2-FLAIR, and a voxelwise reference segmentation. The Additional Training Data (271 examinations from 118 patients) was not used to train the segmentation ensemble; the 188-examination validation release was excluded because it lacked publicly available reference segmentations. Longitudinal MRI scans from LUMIERE and Burdenko underwent the same preprocessing pipeline. Raw data were converted from DICOM to NIfTI and screened for voxel spacing, image geometry, slice thickness, and image clarity. Retained volumes were rigidly registered to the MNI152 template, skull-stripped, and foreground z-score normalized. Each volume was resampled to an isotropic voxel spacing of 1.25×1.25×1.251.25× 1.25× 1.25 m3 and adjusted to 128×128×128128× 128× 128 voxels through foreground cropping, center padding, or center cropping. These steps yield standardized 3D inputs rather than native-resolution, full-field-of-view MRI, allowing consistent volumetric modeling across examinations with different acquisition geometries. Before spatial representation learning, Burdenko scans were associated with abnormal-region probability maps from the BraTS2024-trained segmentation ensemble, whereas LUMIERE scans used the available automated tumor masks. These lesion priors served only as auxiliary structural guidance. Longitudinal pairs were then constructed independently within each patient. The LUMIERE release includes 91 patients and 638 study dates, of which 599 provide all four MRI sequences. Visits were ordered by encoded follow-up week and acquisition phase and retained when contrast-enhanced T1-weighted, native T1-weighted, T2-weighted, and FLAIR images were all available. Consecutive eligible visits were paired, and the later visit’s seven-category annotation was used as the pretraining target. From the 180-patient Burdenko release, we retained patients with curated SD, PsP, or TP clinical-state labels. Follow-up eligibility began at the first recorded follow-up date, with the topometric MRI date and then the end-of-radiotherapy date used as fallback anchors. Only visits acquired on or after the applicable anchor with contrast-enhanced T1-weighted and FLAIR images were retained. Native T1-weighted and T2-weighted images were included when available, with modality-presence indicators supplied to the model. Patients with fewer than two eligible visits were excluded. The remaining visits were ordered chronologically, and only adjacent visits were paired. For each pair, the inter-scan interval was calculated in weeks as Δt=(dt+1−dt)/7 t=(d_t+1-d_t)/7, where dtd_t and dt+1d_t+1 denote the acquisition dates in days. Each pair inherited the patient’s curated SD, PsP, or TP clinical-state label; thus, eligible pairs from the same patient shared that label. The resulting longitudinal pair construction and downstream prediction setup are illustrated in Figure 5. Figure 5: Longitudinal study setup for post-treatment GBM clinical-state classification. Eligible post-treatment MRI scans and their lesion priors form adjacent-visit pairs separated by the inter-scan interval Δt t. Each pair is used to predict the patient’s curated clinical-state label: stable disease (SD), pseudoprogression (PsP), or true progression (TP). 3.2 Training and Implementation Details Model development proceeded in three stages using PyTorch. The segmentation ensemble was trained on BraTS2024 and frozen; STRIDE was pretrained on LUMIERE paired MRI scans and their inter-scan intervals, and its parameters initialized the Burdenko model for downstream SD, PsP, and TP classification. The optimization configurations are summarized in Table 2. A weight decay of 1×10−41× 10^-4 was used throughout training, and gradients were clipped to a maximum ℓ2 _2 norm of 1.0. Table 2: Training configurations for segmentation training, longitudinal pretraining, and downstream clinical-state classification. Stage Optimizer Learning rate Batch size Epochs BraTS2024 segmentation AdamW 1×10−31× 10^-3 2 200 LUMIERE pretraining AdamW 1×10−41× 10^-4 8 300 Burdenko fine-tuning AdamW 8×10−58× 10^-5 8 200 Data augmentation was applied only to training samples. Each spatial axis was independently flipped with probability 0.5, with the same spatial transformation applied jointly to both scans in each longitudinal pair, all MRI channels, and their corresponding lesion masks. Intensity augmentation included random scaling a∼(0.9,1.1)a (0.9,1.1), random shifting b∼(−0.1,0.1)b (-0.1,0.1), and additive Gaussian noise ϵ∼(0,0.032) ε (0,0.03^2). Validation and test samples were not augmented. Class imbalance was addressed through the class-weighted cross-entropy loss in Eq. 20. Class weights were calculated from normalized inverse class frequencies within each training partition and scaled to have a mean of one. For the transition-related objectives, the pretraining coefficients were set to λtr=0.25 _tr=0.25 and λdir=0.10 _dir=0.10, while the downstream coefficients were βtr=0.15 _tr=0.15 and βdir=0.05 _dir=0.05. For downstream classification, model checkpoints were selected according to the highest macro-averaged validation ROC–AUC. During Burdenko fine-tuning, the first 10 epochs were excluded from checkpoint selection. The selected checkpoint was evaluated on the held-out test fold, which was never used for optimization or checkpoint selection. The random seed was fixed at 42 for patient partitioning, data-loader initialization, and model training. During downstream transfer, the pretrained backbone was frozen for the first five epochs and subsequently optimized jointly with the temporal and classification modules. Additional implementation settings included a stem width of 12, 24 hidden channels in the multi-scale localization branches, a drop-path probability of 0.15, an input-adaptation dropout of 0.08, and a temporal-dynamics dropout of 0.10. The temporal scaling constant in Eq. 11 was set to τ=52τ=52 weeks. Eight periodic frequencies were used for temporal encoding and projected to a 64-dimensional temporal embedding. The pair-context representation used an output dimension of 192 and a hidden dimension of 384. 3.3 Evaluation Protocol and Metrics For lesion segmentation, the labeled BraTS2024 examinations were assessed using five examination-level folds supplied with the segmentation ensemble: four for training and one for validation in each iteration, without an independent test set. Among 484 patients with multiple examinations, 415 (85.7%) had scans in multiple folds; thus, the reported segmentation metrics are not patient-independent. To evaluate downstream classification while preventing patient-level information leakage, performance on Burdenko was assessed using patient-level stratified ten-fold cross-validation. All MRI scans and longitudinal pairs derived from the same patient were assigned to the same fold. In each cross-validation iteration, one fold was reserved for final testing, one fold was used for validation and model selection, and the remaining eight folds were used for training. Longitudinal modeling strategies were compared using identical patient-level partitions and evaluation protocols. The evaluated strategies included direct feature fusion, Siamese feature comparison, LSTM-based temporal aggregation, Temporal Transformer modeling, and STRIDE. Whenever applicable, the same visit-level spatial encoder, lesion-prior guidance, data augmentation, and optimization settings were maintained so that the comparison primarily reflected differences in longitudinal modeling strategy. For three-class post-treatment clinical-state classification, macro-averaged one-vs-rest ROC–AUC was used as the primary evaluation metric. Accuracy and macro F1-score were additionally reported. For lesion segmentation, Dice coefficient and Hausdorff distance at the 95th percentile (HD95) were used to evaluate volumetric overlap and boundary agreement, respectively. All classification metrics were calculated on held-out test samples within the patient-level cross-validation procedure. Classification results are reported as mean ± standard deviation across the ten held-out test folds. A pooled out-of-fold macro ROC–AUC was also computed once after aggregating the predictions from all ten test folds. For comparisons with longitudinal baselines, the ten test folds were matched by patient-level partition. Within each fold, the baseline macro ROC–AUC was subtracted from that of STRIDE. The mean difference and its two-sided 95% confidence interval were calculated from these ten paired differences using the Student t distribution with nine degrees of freedom. Two-sided paired t tests were performed on the same differences, and the resulting four p values were adjusted using Holm’s method. Because the training sets overlapped across folds, these comparisons were interpreted as exploratory within-cohort analyses. 4 Results 4.1 Segmentation Performance and Effect of Lesion-Prior Guidance Figure 6: Segmentation performance and downstream effect of lesion-prior guidance. (a) Dice scores for whole tumor (WT), tumor core (TC), and enhancing tumor (ET) on BraTS2024. (b) HD95 values for the corresponding tumor regions, where lower values indicate better boundary agreement. (c) Macro ROC–AUC comparison with and without segmentation-derived lesion priors across different spatial backbones. Error bars denote standard deviation across the ten held-out test folds where applicable. To evaluate lesion localization, the AWHT-based segmentation network was compared with representative segmentation backbones on BraTS2024. As shown in Fig. 6(a,b), AWHT achieved the highest Dice scores and the lowest HD95 values across whole tumor (WT), tumor core (TC), and enhancing tumor (ET). Specifically, Dice scores reached 0.897, 0.761, and 0.798 for WT, TC, and ET, respectively, while the corresponding HD95 values were 3.250, 3.860, and 5.844 m. Complete numerical comparisons are provided in Table A.1. We next examined whether these segmentation-derived priors translated into improved clinical-state classification. As shown in Fig. 6(c), lesion-prior guidance increased macro ROC–AUC across all evaluated architectures. Averaged across models, macro ROC–AUC increased from 0.613 to 0.686. Average accuracy increased from 0.645 to 0.714, while macro F1 increased from 0.609 to 0.668. Complete results for all three metrics are reported in Table A.2. 4.2 Longitudinal Pretraining and Benefits of Longitudinal Input Figure 7: Benefits of longitudinal learning for downstream post-treatment GBM clinical-state classification. (a) Comparison between random initialization and LUMIERE longitudinal pretraining. (b–d) Comparison between single-timepoint and paired longitudinal inputs for macro ROC–AUC, accuracy, and macro F1, respectively. Error bars denote standard deviation across the ten held-out test folds. STRIDE was pretrained on the 499 longitudinal LUMIERE pairs and subsequently adapted to Burdenko, with random initialization used as the reference under identical downstream training settings. As shown in Fig. 7(a), longitudinal pretraining increased macro ROC–AUC from 0.626 to 0.816, accuracy from 0.601 to 0.845, and macro F1 from 0.634 to 0.796. Complete mean ± standard deviation results are provided in Table A.3. We then compared single-timepoint and paired longitudinal configurations across multiple spatial backbones under the same evaluation setting. As shown in Fig. 7(b–d), longitudinal input improved performance across all evaluated architectures. Averaged across models, macro ROC–AUC increased from 0.543 to 0.706, accuracy from 0.676 to 0.767, and macro F1 from 0.542 to 0.684. Complete numerical results are reported in Table A.4. 4.3 Comparison with Longitudinal Modeling Strategies Figure 8: Comparison of longitudinal modeling strategies for three-class post-treatment GBM clinical-state classification. Performance is reported using macro ROC–AUC, accuracy, and macro F1. Error bars denote standard deviation across the ten held-out test folds. STRIDE was compared with direct feature fusion, Siamese feature comparison, recurrent temporal aggregation, and attention-based temporal modeling. As shown in Fig. 8, STRIDE achieved the highest mean performance across all three evaluation metrics, reaching a fold-wise mean macro ROC–AUC of 0.816 ± 0.083, accuracy of 0.845 ± 0.049, and macro F1 of 0.796 ± 0.072. Complete numerical results are provided in Table A.5. The Temporal Transformer was the closest competitor, with a fold-wise mean macro ROC–AUC of 0.768, accuracy of 0.813, and macro F1 of 0.760 (Table A.5). The mean within-fold macro ROC–AUC difference between STRIDE and this baseline was 0.048. Paired comparisons with all four baselines are reported in Table A.7. Pooled out-of-fold predictions yielded a macro ROC–AUC of 0.85 (Fig. A.1), compared with the mean fold-wise macro ROC–AUC of 0.816 reported above. The same figure shows class-wise precision–recall curves and a normalized confusion matrix for SD, PsP, and TP. 4.4 Component Analysis of STRIDE Figure 9: Cumulative component analysis of STRIDE. Performance changes are shown for (a) macro ROC–AUC, (b) accuracy, and (c) macro F1 as interval conditioning, latent transition modeling, directional consistency, and observed–transition fusion are progressively introduced. Error bars denote standard deviation across the ten held-out test folds. A cumulative ablation progressively introduced interval conditioning, gated latent transition modeling, latent directional consistency regularization, and observed–transition fusion. As shown in Fig. 9, performance improved progressively as additional components were introduced. Interval conditioning increased macro ROC–AUC from 0.678 to 0.699. Adding the gated latent transition further increased ROC–AUC to 0.741 and macro F1 from 0.654 to 0.726. Adding latent directional consistency increased macro ROC–AUC to 0.760, and the complete configuration with observed–transition fusion achieved the highest performance, reaching 0.816 macro ROC–AUC, 0.845 accuracy, and 0.796 macro F1. The complete mean ± standard deviation results and exact component configuration for each ablation setting are provided in Table A.6. Figure 10: Qualitative visualization of representative longitudinal GBM cases from true progression (TP), pseudoprogression (PsP), and stable disease (SD). For each case, previous and current contrast-enhanced T1-weighted MRI scans are shown together with segmentation-derived lesion priors, SoftGate activation maps, the inter-scan interval, class probabilities, and the final predicted label. 4.5 Qualitative Analysis of Lesion-aware Longitudinal Representations Figure 10 presents lesion priors and SoftGate activation maps for representative TP, PsP, and SD cases. Across the representative cases, the lesion priors identify tumor-related regions in both examinations, while SoftGate emphasizes these regions without entirely suppressing surrounding tissue. The spatial emphasis also changes between longitudinal visits in accordance with changes in visible lesion distribution rather than remaining fixed across examinations. 5 Discussion The improvements associated with lesion-prior guidance and paired MRI input suggest complementary contributions from spatial representation and longitudinal information. Previous work has examined GBM follow-up MRI at individual timepoints (Guo and Mirzaei, 2026) and longitudinal MRI in high-grade glioma surveillance (Siakallis et al., 2021). In the present cohort, combining scan-specific spatial guidance with longitudinal modeling helped distinguish SD, PsP, and TP. The qualitative findings are consistent with the intended role of SoftGate: emphasizing lesion-related regions while retaining surrounding context. Variable follow-up timing is a broader challenge in irregularly sampled longitudinal modeling (Chen et al., 2018; Rubanova et al., 2019). The cumulative component analysis is consistent with improved performance when interval conditioning and estimated follow-up information are added alongside the observed follow-up representation. Because the components were added in sequence, their independent effects cannot be determined from this ablation. STRIDE also achieved a higher mean fold-wise macro ROC–AUC than the closest Temporal Transformer baseline, with paired comparisons reported in Table A.7. These comparisons reflect performance within the Burdenko cohort. The downstream evaluation used patient-level stratified ten-fold cross-validation within a single retrospective Burdenko cohort. Keeping all scans and pairs from each patient in one fold limited patient-level leakage, but all eligible pairs from that patient shared a curated clinical-state label. The resulting pair-level predictions do not establish visit-specific diagnostic accuracy. Moreover, the BraTS2024 segmentation evaluation used examination-level folds with patient overlap between training and validation, which may have led to optimistic segmentation performance estimates. The quality of the segmentation network’s generated priors on Burdenko was not quantitatively assessed. Independent downstream validation and assessment of lesion-prior quality on post-treatment scans are needed to establish generalizability and clinical utility. 6 Conclusion We presented STRIDE, a longitudinal Transformer framework for three-class post-treatment GBM clinical-state classification from paired MRI scans and their inter-scan interval. STRIDE combines lesion-prior-guided spatial representation, time-conditioned latent transition, and observed–transition fusion in a unified framework. On Burdenko, it achieved the highest mean performance among the evaluated longitudinal strategies. These results support further evaluation of interval-conditioned disease representations on independent post-treatment GBM cohorts. Data and code availability The BraTS2024, LUMIERE, and Burdenko Glioblastoma Progression datasets are publicly available through the BraTS challenge platform, the corresponding LUMIERE data repository, and The Cancer Imaging Archive (TCIA), respectively. The code used to implement the proposed STRIDE framework will be made publicly available on GitHub upon publication at https://github.com/AdrianGuo2484/STRIDE. Acknowledgments Computational resources were provided by the Ohio Supercomputer Center (OSC). Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. CRediT authorship contribution statement W.G.: Conceptualization, Methodology, Software, Investigation, Formal analysis, Writing – original draft. C.Y.: Methodology. C.Y. also provided technical guidance and assisted with interpretation of experimental results. W.C.: Supervision, Writing – review & editing. G.M.: Methodology, Supervision, Writing – review & editing. P.Z.: Methodology, Writing – review & editing. P.G.: Methodology, Writing – review & editing. P.G. also contributed to clinical interpretation of the results. Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this work, the authors used OpenAI ChatGPT to assist with language editing. The authors reviewed and edited the resulting content and take full responsibility for the content of the published article. Appendix A Additional Quantitative Results This appendix provides the complete numerical results corresponding to the quantitative comparisons presented in the main manuscript. Tables A.1–A.7 report the detailed performance values supporting the segmentation evaluation, lesion-prior guidance analysis, longitudinal pretraining experiment, single-versus-longitudinal comparison, longitudinal modeling strategy comparison, STRIDE component analysis, and paired comparison with the evaluated baselines, respectively. Figure A.1 additionally presents class-wise discrimination and prediction patterns of the final STRIDE model. Table A.1: Complete segmentation performance comparison on BraTS2024. Dice (↑ ) and HD95 (↓ , m) are reported for whole tumor (WT), tumor core (TC), and enhancing tumor (ET). Backbone Dice ↑ HD95 ↓ WT TC ET WT TC ET CNN-UNet 0.801 0.627 0.602 4.506 8.483 10.613 MedNeXt 0.885 0.706 0.786 4.390 4.964 7.167 SwinUNet 0.781 0.738 0.727 4.870 4.857 7.422 AWHT Encoder (Ours) 0.897 0.761 0.798 3.250 3.860 5.844 Table A.2: Complete comparison of segmentation-derived lesion-prior guidance for three-class post-treatment clinical-state classification. Macro-averaged one-vs-rest ROC–AUC, accuracy, and macro F1 are reported as mean ± standard deviation across the ten held-out test folds. The Average row reports performance averaged across the evaluated architectures. Macro ROC–AUC ↑ Accuracy ↑ Macro F1 ↑ Backbone w/o Prior w/ Prior w/o Prior w/ Prior w/o Prior w/ Prior CNN 0.466 ± 0.197 0.563 ± 0.155 0.615 ± 0.167 0.685 ± 0.136 0.590 ± 0.141 0.553 ± 0.104 ViT 0.679 ± 0.138 0.754 ± 0.075 0.634 ± 0.044 0.767 ± 0.096 0.592 ± 0.068 0.720 ± 0.142 UNet 0.623 ± 0.071 0.675 ± 0.068 0.638 ± 0.102 0.615 ± 0.034 0.637 ± 0.141 0.654 ± 0.155 SwinUNet 0.611 ± 0.072 0.667 ± 0.049 0.654 ± 0.088 0.662 ± 0.064 0.612 ± 0.058 0.633 ± 0.095 MedNeXt 0.600 ± 0.137 0.633 ± 0.094 0.646 ± 0.086 0.653 ± 0.050 0.626 ± 0.139 0.654 ± 0.184 Swin Transformer 0.631 ± 0.051 0.693 ± 0.038 0.667 ± 0.037 0.769 ± 0.049 0.575 ± 0.076 0.667 ± 0.056 STRIDE 0.678 ± 0.116 0.816 ± 0.083 0.659 ± 0.171 0.845 ± 0.049 0.634 ± 0.077 0.796 ± 0.072 Average 0.613 0.686 0.645 0.714 0.609 0.668 Table A.3: Effect of longitudinal pretraining on downstream three-class post-treatment GBM clinical-state classification. Results are reported as mean ± standard deviation across the ten held-out test folds. Initialization Macro ROC–AUC ↑ Accuracy ↑ Macro F1 ↑ Random initialization 0.626 ± 0.192 0.601 ± 0.221 0.634 ± 0.174 LUMIERE pretraining 0.816 ± 0.083 0.845 ± 0.049 0.796 ± 0.072 Table A.4: Complete comparison between single-timepoint and longitudinal modeling for three-class post-treatment classification. Macro-averaged one-vs-rest ROC–AUC, accuracy, and macro F1 are reported as mean ± standard deviation across the ten held-out test folds. The Average row reports performance averaged across the evaluated architectures. Macro ROC–AUC ↑ Accuracy ↑ Macro F1 ↑ Model Single Longitudinal Single Longitudinal Single Longitudinal CNN 0.465 ± 0.012 0.563 ± 0.155 0.669 ± 0.058 0.685 ± 0.236 0.461 ± 0.046 0.553 ± 0.104 ViT 0.528 ± 0.012 0.754 ± 0.075 0.621 ± 0.015 0.767 ± 0.096 0.476 ± 0.071 0.720 ± 0.142 Swin Transformer 0.481 ± 0.039 0.693 ± 0.038 0.701 ± 0.054 0.769 ± 0.049 0.484 ± 0.039 0.667 ± 0.056 STRIDE 0.699 ± 0.046 0.816 ± 0.083 0.712 ± 0.058 0.845 ± 0.049 0.748 ± 0.102 0.796 ± 0.072 Average 0.543 0.706 0.676 0.767 0.542 0.684 Table A.5: Complete comparison with longitudinal modeling strategies for three-class post-treatment classification. Macro-averaged one-vs-rest ROC–AUC, accuracy, and macro F1 are reported as mean ± standard deviation across the ten held-out test folds. Model Temporal strategy Macro ROC–AUC ↑ Accuracy ↑ Macro F1 ↑ Direct Feature Fusion Concatenation / difference 0.681 ± 0.095 0.695 ± 0.082 0.628 ± 0.125 Siamese Cross-visit feature comparison 0.739 ± 0.062 0.788 ± 0.105 0.754 ± 0.094 LSTM Recurrent temporal aggregation 0.656 ± 0.057 0.698 ± 0.099 0.631 ± 0.077 Temporal Transformer Attention-based temporal modeling 0.768 ± 0.049 0.813 ± 0.069 0.760 ± 0.081 STRIDE Time-conditioned latent transition 0.816 ± 0.083 0.845 ± 0.049 0.796 ± 0.072 Table A.6: Complete cumulative ablation study of STRIDE for three-class post-treatment classification. Macro-averaged one-vs-rest ROC–AUC, accuracy, and macro F1 are reported as mean ± standard deviation across the ten held-out test folds. Dir. Loss denotes latent directional consistency loss and Obs.–Trans. denotes observed–transition fusion. Setting Interval Transition Dir. Loss Obs.–Trans. ROC–AUC ↑ Accuracy ↑ Macro F1 ↑ Baseline Fusion 0.678 ± 0.095 0.695 ± 0.082 0.628 ± 0.125 + Time Interval ✓ 0.699 ± 0.033 0.726 ± 0.081 0.654 ± 0.148 + Latent Transition ✓ ✓ 0.741 ± 0.092 0.769 ± 0.108 0.726 ± 0.039 + Directional Consistency ✓ ✓ ✓ 0.760 ± 0.077 0.783 ± 0.101 0.745 ± 0.062 Full STRIDE ✓ ✓ ✓ ✓ 0.816 ± 0.083 0.845 ± 0.049 0.796 ± 0.072 Table A.7: Paired comparison of fold-wise macro-averaged one-vs-rest ROC–AUC on the Burdenko cohort. Performance is reported as mean ± standard deviation across the ten patient-level test folds. Δ is the mean within-fold difference between STRIDE and the indicated baseline. Adjusted p values account for the four baseline comparisons. Model Macro ROC–AUC Δ ROC–AUC 95% CI for Δ Holm p Direct Feature Fusion 0.681 ± 0.095 0.1350.135 [0.121, 0.149] 1.39×10−81.39× 10^-8 Siamese 0.739 ± 0.062 0.0770.077 [0.050, 0.103] 2.05×10−42.05× 10^-4 LSTM 0.656 ± 0.057 0.1600.160 [0.134, 0.186] 6.46×10−76.46× 10^-7 Temporal Transformer 0.768 ± 0.049 0.0480.048 [0.023, 0.072] 0.00144 STRIDE 0.816 ± 0.083 — — — Figure A.1: Class-wise performance of STRIDE based on pooled out-of-fold predictions from the patient-level ten-fold cross-validation. (a) One-vs-rest receiver operating characteristic (ROC) curves for stable disease (SD), pseudoprogression (PsP), and true progression (TP). After aggregating the held-out predictions from all ten test folds, the pooled out-of-fold macro ROC–AUC was 0.85. (b) Precision–recall curves computed from the pooled out-of-fold predictions for the three clinical-state classes. (c) Normalized confusion matrix computed from the pooled out-of-fold predictions. The pooled ROC–AUC reported here differs from the fold-wise macro ROC–AUC of 0.816±0.0830.816± 0.083 reported in the main manuscript because the former is calculated once after aggregating all held-out predictions, whereas the latter represents the mean ± standard deviation of ROC–AUC values calculated independently across the ten test folds. References Aghajanian et al. (2025) S. Aghajanian, F. Mohammadifard, I. Mohammadi, S. Rajai Firouzabadi, A. Baradaran Bagheri, E. Moases Ghaffary, and O. Mirmosayyeb Longitudinal structural mri-based deep learning and radiomics features for predicting alzheimer’s disease progression. Alzheimer’s Research & Therapy 17 (1), p. 182. Cited by: §1. Akbari et al. (2020) H. Akbari, S. Rathore, S. Bakas, M. P. Nasrallah, G. Shukla, E. Mamourian, M. Rozycki, S. J. Bagley, J. D. Rudie, A. E. Flanders, et al. Histopathology-validated machine learning radiographic biomarker for noninvasive discrimination between true progression and pseudo-progression in glioblastoma. Cancer 126 (11), p. 2625–2636. Cited by: §1. Brandes et al. (2008) A. A. Brandes, E. Franceschi, A. Tosoni, V. Blatt, A. Pession, G. Tallini, R. Bertorelle, S. Bartolini, F. Calbucci, A. Andreoli, et al. MGMT promoter methylation status can predict the incidence and outcome of pseudoprogression after concomitant radiochemotherapy in newly diagnosed glioblastoma patients. Journal of Clinical Oncology 26 (13), p. 2192–2197. Cited by: §1. Che et al. (2018) Z. Che, S. Purushotham, K. Cho, D. Sontag, and Y. Liu Recurrent neural networks for multivariate time series with missing values. Scientific reports 8 (1), p. 6085. Cited by: §1. Chen et al. (2018) R. T. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud Neural ordinary differential equations. Advances in neural information processing systems 31. Cited by: §1, §5. Cui et al. (2019) R. Cui, M. Liu, A. D. N. Initiative, et al. RNN-based longitudinal analysis for diagnosis of alzheimer’s disease. Computerized Medical Imaging and Graphics 73, p. 1–10. Cited by: §1, §1. De Brouwer et al. (2019) E. De Brouwer, J. Simm, A. Arany, and Y. Moreau GRU-ode-bayes: continuous modeling of sporadically-observed time series. Advances in neural information processing systems 32. Cited by: §1. de Verdier et al. (2024) M. C. de Verdier, R. Saluja, L. Gagnon, D. LaBella, U. Baid, N. H. Tahon, M. Foltyn-Dumitru, J. Zhang, M. Alafif, S. Baig, et al. The 2024 brain tumor segmentation (brats) challenge: glioma segmentation on post-treatment mri. arXiv preprint arXiv:2405.18368. Cited by: §3.1. Elshafeey et al. (2019) N. Elshafeey, A. Kotrotsou, A. Hassan, N. Elshafei, I. Hassan, S. Ahmed, S. Abrol, A. Agarwal, K. El Salek, S. Bergamaschi, et al. Multicenter study demonstrates radiomic features derived from magnetic resonance perfusion images identify pseudoprogression in glioblastoma. Nature communications 10 (1), p. 3170. Cited by: §1. Fang et al. (2019) L. Fang, C. Wang, S. Li, H. Rabbani, X. Chen, and Z. Liu Attention to lesion: lesion-aware convolutional neural network for retinal optical coherence tomography image classification. IEEE transactions on medical imaging 38 (8), p. 1959–1970. Cited by: §1. Gao et al. (2026) Y. Gao, Y. Zhang, C. Wang, J. Liu, M. Varma, J. Delbrouck, A. Chaudhari, and C. Langlotz Learning generalizable 3d medical image representations from mask-guided self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 13744–13754. Cited by: §1. Ghazi et al. (2019) M. M. Ghazi, M. Nielsen, A. Pai, M. J. Cardoso, M. Modat, S. Ourselin, L. Sørensen, A. D. N. Initiative, et al. Training recurrent neural networks robust to incomplete data: application to alzheimer’s disease progression modeling. Medical image analysis 53, p. 39–46. Cited by: §1. Gomaa et al. (2025) A. Gomaa, Y. Huang, P. Stephan, K. Breininger, B. Frey, A. Dörfler, O. Schnell, D. Delev, R. Coras, A. Donaubauer, et al. A self-supervised multimodal deep learning approach to differentiate post-radiotherapy progression from pseudoprogression in glioblastoma. Scientific Reports 15 (1), p. 17133. Cited by: §1, §1, §1. Guo and Mirzaei (2026) W. Guo and G. Mirzaei Timepoint-specific benchmarking of deep learning models for glioblastoma follow-up mri. Cancers 18 (1), p. 36. Cited by: §1, §1, §5. Hatamizadeh et al. (2022a) A. Hatamizadeh, V. Nath, Y. Tang, D. Yang, H. R. Roth, and D. Xu Swin UNETR: swin transformers for semantic segmentation of brain tumors in mri images. In Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. BrainLes 2021, A. Crimi and S. Bakas (Eds.), Lecture Notes in Computer Science, Vol. 12962, p. 352–364. External Links: Document, Link Cited by: §1. Hatamizadeh et al. (2022b) A. Hatamizadeh, Y. Tang, V. Nath, D. Yang, A. Myronenko, B. Landman, H. R. Roth, and D. Xu Unetr: transformers for 3d medical image segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, p. 574–584. Cited by: §1. Huang et al. (2026) X. Huang, Z. Xu, Y. Zhao, Y. Wang, Y. Liu, W. Hu, K. Zhao, L. Yao, J. He, Y. Yu, et al. Longitudinal mri-based deep learning model for predicting pathological complete response in breast cancer: a multicenter, retrospective cohort study. npj Precision Oncology. Cited by: §1. Huo et al. (2025) Z. F. Huo, J. Fries, A. Lozano, J. M. J. Valanarasu, E. Steinberg, L. Blankemeier, A. Chaudhari, C. Langlotz, and N. Shah Time-to-event pretraining for 3d medical imaging. In International Conference on Learning Representations, Vol. 2025, p. 84072–84108. Cited by: §1. Isensee et al. (2021) F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein NnU-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18 (2), p. 203–211. Cited by: §1. Jang et al. (2018) B. Jang, S. H. Jeon, I. H. Kim, and I. A. Kim Prediction of pseudoprogression versus progression using machine learning algorithm in glioblastoma. Scientific reports 8 (1), p. 12516. Cited by: §1. Kim et al. (2019) J. Y. Kim, J. E. Park, Y. Jo, W. H. Shim, S. J. Nam, J. H. Kim, R. Yoo, S. H. Choi, and H. S. Kim Incorporating diffusion-and perfusion-weighted mri into a radiomics model improves diagnostic performance for pseudoprogression in glioblastoma patients. Neuro-oncology 21 (3), p. 404–414. Cited by: §1. Kotochinsky et al. (2026) M. Kotochinsky, Y. P. Silva, A. Goyal, P. Łajczak, W. Rios-Garcia, R. Suruagy-Motta, I. N. S. Petry, T. P. Fagundes, and R. Bertani Machine learning models for predicting pseudoprogression in glioblastoma: a systematic review and diagnostic meta-analysis. Neuro-Oncology Practice, p. npag055. Cited by: §1. Li et al. (2025) J. Li, R. Liu, Y. Xing, Q. Yin, and Q. Su A deep learning model for distinguishing pseudoprogression and tumor progression in glioblastoma based on pre-and post-operative contrast-enhanced t1 imaging. Radiotherapy and Oncology, p. 111111. Cited by: §1. Li et al. (2020) M. D. Li, K. Chang, B. Bearce, C. Y. Chang, A. J. Huang, J. P. Campbell, J. M. Brown, P. Singh, K. V. Hoebel, D. Erdoğmuş, et al. Siamese neural networks for continuous disease severity evaluation and change detection in medical imaging. NPJ digital medicine 3 (1), p. 48. Cited by: §1, §1. Liu et al. (2025) C. Liu, K. Xu, L. L. Shen, G. Huguet, Z. Wang, A. Tong, D. Bzdok, J. Stewart, J. C. Wang, L. V. Del Priore, et al. Imageflownet: forecasting multiscale image-level trajectories of disease progression with irregularly-sampled longitudinal medical images. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 1–5. Cited by: §1. Mehta et al. (2018) S. Mehta, E. Mercan, J. Bartlett, D. Weaver, J. G. Elmore, and L. Shapiro Y-net: joint segmentation and classification for diagnosis of breast biopsy images. In International conference on medical image computing and computer-assisted intervention, p. 893–901. Cited by: §1. Moassefi et al. (2022) M. Moassefi, S. Faghani, G. M. Conte, R. O. Kowalchuk, S. Vahdati, D. J. Crompton, C. Perez-Vega, R. A. D. Cabreja, S. A. Vora, A. Quiñones-Hinojosa, et al. A deep learning model for discriminating true progression from pseudoprogression in glioblastoma patients. Journal of neuro-oncology 159 (2), p. 447–455. Cited by: §1. Mouadden et al. (2025) N. Mouadden, O. Laousy, R. Marini, V. Ong, M. Revel, G. Chassagnon, S. Christodoulidis, and M. Vakalopoulou Conditional latent diffusion models for irregularly spaced longitudinal radiological data. In International Conference on Medical Image Computing and Computer-Assisted Intervention, p. 106–115. Cited by: §1. Nguyen et al. (2023) H. H. Nguyen, M. B. Blaschko, S. Saarakkala, and A. Tiulpin Clinically-inspired multi-agent transformers for disease trajectory forecasting from multimodal data. IEEE transactions on medical imaging 43 (1), p. 529–541. Cited by: §1. Okada et al. (2018) H. Okada, K. M. Downey, and D. A. Reardon Immunotherapy response assessment in neuro-oncology (irano). In Handbook of Brain Tumor Chemotherapy, Molecular Therapeutics, and Immunotherapy, p. 761–766. Cited by: §1. Okada et al. (2015) H. Okada, M. Weller, R. Huang, G. Finocchiaro, M. R. Gilbert, W. Wick, B. M. Ellingson, N. Hashimoto, I. F. Pollack, A. A. Brandes, et al. Immunotherapy response assessment in neuro-oncology: a report of the rano working group. The Lancet Oncology 16 (15), p. e534–e542. Cited by: §1. Palavani et al. (2025) L. B. Palavani, B. V. Nogueira, L. P. Mitre, H. Chen, G. C. Müller, M. Vilardo, V. G. Pereira, L. F. F. Paleare, F. V. Ribeiro, A. A. S. Araujo, et al. Artificial intelligence algorithms for differentiating pseudoprogression from true progression in high-grade gliomas: a systematic review and meta-analysis. Neurosurgical Review 48 (1), p. 591. Cited by: §1. Ronneberger et al. (2015) O. Ronneberger, P. Fischer, and T. Brox U-net: convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, p. 234–241. Cited by: §1. Rubanova et al. (2019) Y. Rubanova, R. T. Chen, and D. K. Duvenaud Latent ordinary differential equations for irregularly-sampled time series. Advances in neural information processing systems 32. Cited by: §1, §5. Siakallis et al. (2021) L. Siakallis, C. H. Sudre, P. Mulholland, N. Fersht, J. Rees, L. Topff, S. Thust, R. Jager, M. J. Cardoso, J. Panovska-Griffiths, et al. Longitudinal structural and perfusion mri enhanced by machine learning outperforms standalone modalities and radiological expertise in high-grade glioma surveillance. Neuroradiology 63 (12), p. 2047–2056. Cited by: §1, §5. Stupp et al. (2005) R. Stupp, W. P. Mason, M. J. Van Den Bent, M. Weller, B. Fisher, M. J. Taphoorn, K. Belanger, A. A. Brandes, C. Marosi, U. Bogdahn, et al. Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. New England journal of medicine 352 (10), p. 987–996. Cited by: §1. Sun et al. (2024) Y. Sun, K. Li, D. Chen, Y. Hu, and S. Zhang LOMIA-t: a transformer-based longitudinal medical image analysis framework for predicting treatment response of esophageal cancer. In International Conference on Medical Image Computing and Computer-Assisted Intervention, p. 426–436. Cited by: §1, §1. Suter et al. (2022) Y. Suter, U. Knecht, W. Valenzuela, M. Notter, E. Hewer, P. Schucht, R. Wiest, and M. Reyes The lumiere dataset: longitudinal glioblastoma mri with expert rano evaluation. Scientific data 9 (1), p. 768. Cited by: §1, §3.1. Taal et al. (2008) W. Taal, D. Brandsma, H. G. de Bruin, J. E. Bromberg, A. T. Swaak-Kragten, P. A. Sillevis Smitt, C. A. van Es, and M. J. van den Bent Incidence of early pseudo-progression in a cohort of malignant glioma patients treated with chemoirradiation with temozolomide. Cancer 113 (2), p. 405–410. Cited by: §1. Tang et al. (2022) Y. Tang, D. Yang, W. Li, H. R. Roth, B. Landman, D. Xu, V. Nath, and A. Hatamizadeh Self-supervised pre-training of swin transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 20730–20740. Cited by: §1. Tarek et al. (2026) A. Tarek, H. Saberr, H. Elghonemy, Y. Afify, T. Basha, O. S. Bhatti, A. M. Selim, H. M. T. Alam, and D. Sonntag TRACE: a concept bottleneck model for longitudinal 3d glioblastoma response assessment. arXiv preprint arXiv:2606.30313. Cited by: §1, §1. Turk et al. (2022) S. Turk, N. C. Wang, O. Kitis, S. Mohammed, T. Ma, R. Lobo, J. Kim, S. Camelo-Piragua, T. D. Johnson, M. M. Kim, et al. Comparative study of radiologists vs machine learning in differentiating biopsy-proven pseudoprogression and true progression in diffuse gliomas. Neuroscience Informatics 2 (3), p. 100088. Cited by: §1. Wang et al. (2024) X. Wang, T. Tan, Y. Gao, E. Marcus, L. Han, A. Portaluri, T. Zhang, C. Lu, X. Liang, R. Beets-Tan, et al. Ordinal learning: longitudinal attention alignment model for predicting time to future breast cancer events from mammograms. In International Conference on Medical Image Computing and Computer-Assisted Intervention, p. 155–165. Cited by: §1. Wen et al. (2010) P. Y. Wen, D. R. Macdonald, D. A. Reardon, T. F. Cloughesy, A. G. Sorensen, E. Galanis, J. DeGroot, W. Wick, M. R. Gilbert, A. B. Lassman, et al. Updated response assessment criteria for high-grade gliomas: response assessment in neuro-oncology working group. Journal of clinical oncology 28 (11), p. 1963–1972. Cited by: §1. Wen et al. (2023) P. Y. Wen, M. Van Den Bent, G. Youssef, T. F. Cloughesy, B. M. Ellingson, M. Weller, E. Galanis, D. P. Barboriak, J. De Groot, M. R. Gilbert, et al. RANO 2.0: update to the response assessment in neuro-oncology criteria for high-and low-grade gliomas in adults. Journal of Clinical Oncology 41 (33), p. 5187–5199. Cited by: §1. Young et al. (2023) J. S. Young, N. Al-Adli, K. Scotford, S. Cha, and M. S. Berger Pseudoprogression versus true progression in glioblastoma: what neurosurgeons need to know. Journal of neurosurgery 139 (3), p. 748–759. Cited by: §1. Zhang et al. (2025) S. Zhang, Y. Sun, Y. Ao, X. Zhang, R. Yang, J. Xu, Z. Ai, H. Zhang, X. Yang, Y. Xu, et al. GLOMIA-pro: a generalizable longitudinal medical image analysis framework for disease progression prediction. arXiv preprint arXiv:2507.12500. Cited by: §1, §1. Zheng et al. (2025) G. Zheng, Y. Lu, and H. Chen Deep learning based framework for predicting mild cognitive impairment progression in neurology using longitudinal mri. IEEE Access. Cited by: §1. Zheng et al. (2026) Y. Zheng, D. Huang, J. Wei, T. Liu, X. Wu, Y. Feng, C. Chen, and Y. Liu Whole-process evolutionary heterogeneity analysis for glioblastoma radiotherapy response prediction. IEEE Journal of Biomedical and Health Informatics. Cited by: §1, §1. Zhou et al. (2021) Z. Zhou, V. Sodha, J. Pang, M. B. Gotway, and J. Liang Models genesis. Medical image analysis 67, p. 101840. Cited by: §1. Zolotova et al. (2023) S. V. Zolotova, A. V. Golanov, I. N. Pronin, A. V. Dalechina, A. A. Nikolaeva, A. S. Belyashova, D. Y. Usachev, E. A. Kondrateva, P. V. Druzhinina, B. N. Shirokikh, T. N. Saparov, M. G. Belyaev, and A. I. Kurmukov Burdenko’s glioblastoma progression dataset (burdenko-gbm-progression). Note: The Cancer Imaging ArchiveVersion 1 External Links: Document Cited by: §3.1.