STRIDE: Spatial-Temporal Representation for Interval-conditioned Disease Evolution in Longitudinal Glioblastoma MRI

1. Introduction

• Post-treatment appearances can overlap across stable disease, pseudoprogression, and true progression despite different disease trajectories, making their reliable distinction challenging.Distinguishing pseudoprogression from true progression is especially difficult during early post-treatment follow-up.Source: §1 · p. 2
• Longitudinal methods face three gaps: adapting to scan-specific spatial variability, incorporating actual inter-scan intervals, and jointly modeling follow-up state and longitudinal change.A single MRI examination provides only a snapshot and does not directly capture change between visits.Source: §1 · p. 3
• Longitudinal pretraining is used to support representation transfer before downstream adaptation because annotated longitudinal GBM cohorts remain limited.The authors cite prior work on pretraining and transfer learning.Source: §1 · p. 4
• STRIDE predicts SD, PsP, or TP from paired post-treatment MRI scans and their interval, combining lesion-prior-guided spatial representation, time-conditioned latent transition, and observed–transition fusion.The spatial representation emphasizes lesion-related regions while retaining context; transition and fusion model interval-dependent change and complementary state–change information.Source: §1 · pp. 3–4

2.1. Overview of STRIDE Framework

• STRIDE comprises lesion-prior-guided spatial representation, time-conditioned latent transition, and observed–transition fusion for three-class clinical-state classification.The components extract visit-specific spatial representations, characterize interval-dependent change, and integrate observed with transition-estimated longitudinal information.Source: §2.1 · p. 4, Figure 1 · p. 5

2.2. Problem Formulation

• Each sample is a pair of MRI visits and their inter-scan interval, with the patient’s curated SD, PsP, or TP label assigned to the pair.The formulation denotes each visit by its scan and acquisition time and predicts probabilities for the three clinical states.Source: §2.2 · p. 5

2.3. Lesion-prior-guided Spatial Representation

• Post-treatment abnormalities are localized, but surrounding tissue may also inform SD, PsP, and TP discrimination; the method combines structural guidance with soft lesion-aware modulation and adaptive spatial learning.The motivation contrasts treating all locations equally with restricting learning to a hard lesion mask, which removes contextual features.Source: §2.3 · p. 6

2.3.2. Mask-conditioned SoftGate

• SoftGate uses the segmentation prior to modulate visit features, applying lesion-aware emphasis while preserving the original feature representation.The prior is spatially aligned and projected to match the visit-specific feature map; modulation strength α is fixed at 0.5 in all experiments.Source: §2.3 · p. 6, §2.3.2 · p. 7

2.3.1. Segmentation-derived Lesion Prior

• A frozen segmentation network generates a lesion prior for each visit, used only as auxiliary structural guidance for spatial representation learning.The prior is a voxel-wise abnormal-region probability map.Source: §2.3.1 · pp. 6–7

2.3.3. Adaptive-window Hierarchical Transformer

• AWHT estimates an examination-specific continuous spatial window from multi-scale features before hierarchical window-based Transformer encoding.It uses localization branches at scales 1, 2, and 4; the local self-attention window remains fixed at 8 × 8 × 8 tokens.Source: §2.3.3 · pp. 7–8

2.4. Time-conditioned Latent Transition

• STRIDE combines interval encoding, pair-level context modeling, and gated transition estimation to characterize scan relationships over irregular follow-up intervals.The transition uses both observed visit representations to compute pair-level context.Source: §2.4 · p. 9

2.4.1. Inter-scan Interval Encoding

• The inter-scan interval is log-scaled and periodically encoded into an embedding that conditions the subsequent latent transition.The interval embedding is produced by mapping the encoded interval through a multilayer perceptron.Source: §2.4.1 · p. 9

2.4.2. Pair-level Context Modeling

• Pair-level context combines the two observed visit representations, their observed latent change, and an adaptively weighted cross-visit interaction.The interaction supplements additive change, which does not capture interactions between corresponding latent dimensions.Source: §2.4.2 · pp. 9–11

2.4.3. Gated Latent Transition

• The earlier representation, pair-level context, and interval embedding produce a gated, interval-scaled transition estimate of the follow-up representation.The estimate is interpreted as pair-conditioned interval-dependent change, not a future-state prediction from the earlier visit alone.Source: §2.4.3 · p. 11

2.5. Observed–Transition Fusion

• The fusion module integrates observed and transition-estimated follow-up representations, latent changes, pair-level context, and interval embedding through multi-source cross-attention.The sources are projected into a common representation space and aggregated into a unified fused representation.Source: §2.5 · pp. 11–12, Figure 4 · p. 12
• The fused representation is refined by an MLP and passed to a classification head with softmax to predict SD, PsP, or TP.The classification head produces a probability vector.Source: §2.5 · p. 13

2.6. Training Objectives

• The training objective combines class-weighted classification with transition and latent directional consistency losses.The transition loss constrains the magnitude of discrepancy from the observed follow-up representation, while directional consistency encourages latent-space alignment.Source: §2.6 · p. 13
• The class-weighted cross-entropy uses seven classes for LUMIERE pretraining and three for Burdenko downstream adaptation.The class weights correspond to the target classes at each training stage.Source: §2.6 · p. 13

3.1. Datasets and Data Preparation

• BraTS2024 supplied annotated examinations for lesion-prior generation, LUMIERE supplied longitudinal pairs for pretraining, and curated Burdenko visits supplied adjacent pairs for SD, PsP, or TP classification.LUMIERE pairs used the later visit's annotation as the pretraining target; Burdenko pairs inherited the patient's curated clinical-state label.Source: §3.1 · pp. 14–16

3.2. Training and Implementation Details

• Training proceeded from a frozen BraTS2024 segmentation ensemble to LUMIERE longitudinal pretraining and then Burdenko adaptation.During transfer, the pretrained backbone was frozen for the first five epochs before joint optimization with temporal and classification modules.Source: §3.2 · pp. 16–17, Table 2 · p. 17

3.3. Evaluation Protocol and Metrics

• Burdenko classification used patient-level stratified ten-fold cross-validation, whereas BraTS2024 segmentation used examination-level folds without an independent test set.Because training sets overlapped across folds, baseline comparisons were interpreted as exploratory within-cohort analyses.Source: §3.3 · pp. 18–19

4.1. Segmentation Performance and Effect of Lesion-Prior Guidance

• 0.897, 0.761, and 0.798 Dice scores for WT, TC, and ET accompanied the highest Dice and lowest HD95 across those regions for AWHT.With lesion-prior guidance, average macro ROC–AUC increased from 0.613 to 0.686, accuracy from 0.645 to 0.714, and macro F1 from 0.609 to 0.668.Source: §4.1 · pp. 19–20

4.2. Longitudinal Pretraining and Benefits of Longitudinal Input

• 0.626 to 0.816 macro ROC–AUC, 0.601 to 0.845 accuracy, and 0.634 to 0.796 macro F1 followed LUMIERE pretraining; paired inputs also improved all evaluated spatial backbones.Across models, longitudinal input raised macro ROC–AUC from 0.543 to 0.706, accuracy from 0.676 to 0.767, and macro F1 from 0.542 to 0.684.Source: Figure 7 · p. 20, §4.2 · p. 21

4.3. Comparison with Longitudinal Modeling Strategies

• 0.816 ± 0.083 macro ROC–AUC, 0.845 ± 0.049 accuracy, and 0.796 ± 0.072 macro F1 were STRIDE's highest mean results across the evaluated longitudinal strategies.The Temporal Transformer was the closest competitor, with a fold-wise mean macro ROC–AUC of 0.768; STRIDE's mean within-fold macro ROC–AUC difference was 0.048.Source: Figure 8 · p. 21, §4.3 · p. 22
• 0.85 macro ROC–AUC was obtained from pooled out-of-fold predictions, compared with the mean fold-wise macro ROC–AUC of 0.816.The pooled figure also presents class-wise precision–recall curves and a normalized confusion matrix for SD, PsP, and TP.Source: §4.3 · p. 22

4.4. Component Analysis of STRIDE

• 0.816 macro ROC–AUC, 0.845 accuracy, and 0.796 macro F1 marked the complete configuration, after successive component additions improved performance.Interval conditioning raised macro ROC–AUC from 0.678 to 0.699; the gated latent transition raised it to 0.741 and macro F1 from 0.654 to 0.726; directional consistency raised macro ROC–AUC to 0.760.Source: §4.4 · pp. 22–23, Figure 9 · p. 22

4.5. Qualitative Analysis of Lesion-aware Longitudinal Representations

• Representative TP, PsP, and SD cases showed SoftGate emphasizing lesion regions without entirely suppressing surrounding tissue, with emphasis changing across visits.The spatial emphasis varied in accordance with visible lesion-distribution changes.Source: §4.5 · pp. 23–24, Figure 10 · p. 23

• Improvements associated with lesion-prior guidance and paired MRI input suggest complementary contributions from spatial representation and longitudinal information.Qualitative findings are consistent with SoftGate emphasizing lesion-related regions while retaining surrounding context.Source: §5 · p. 24
• STRIDE achieved a higher mean fold-wise macro ROC–AUC than the closest Temporal Transformer baseline in the Burdenko cohort.The component ablation adds components in sequence, so their independent effects cannot be determined.Source: §5 · p. 24
• The single retrospective Burdenko evaluation does not establish visit-specific diagnostic accuracy, and independent validation is needed to establish generalizability and clinical utility.Pairs shared a curated clinical-state label; BraTS2024 segmentation folds had patient overlap, and prior quality on Burdenko was not quantitatively assessed.Source: §5 · p. 24

• On Burdenko, STRIDE achieved the highest mean performance among the evaluated longitudinal strategies.Source: §6 · p. 25
• STRIDE classifies three post-treatment GBM clinical states from paired MRI scans and their inter-scan interval.Source: §6 · p. 25
• The framework combines lesion-prior-guided spatial representation, time-conditioned latent transition, and observed–transition fusion.Source: §6 · p. 25
• The Burdenko results support further evaluation of interval-conditioned disease representations on independent post-treatment GBM cohorts.Source: §6 · p. 25

• W.G. is credited with conceptualization, methodology, software, investigation, formal analysis, and the original draft.Source: CRediT authorship contribution statement · p. 25
• C.Y. contributed methodology, technical guidance, and interpretation of experimental results.Source: CRediT authorship contribution statement · p. 25
• W.C., G.M., P.Z., and P.G. contributed to methodology and manuscript review; W.C. and G.M. also supervised, and P.G. contributed clinical interpretation.Source: CRediT authorship contribution statement · pp. 25–26

• Tables A.1–A.4 provide detailed results for BraTS2024 segmentation, lesion-prior guidance, longitudinal pretraining, and single-timepoint versus longitudinal modeling.Source: Appendix A · p. 27, Table A.1 · p. 27, Table A.2 · p. 27
• Tables A.5–A.7 report comparisons of longitudinal strategies, cumulative STRIDE components, and paired fold-wise macro ROC–AUC against evaluated baselines.Source: Appendix A · p. 27, Table A.5 · p. 28, Table A.6 · p. 28
• Figure A.1 reports a pooled out-of-fold macro ROC–AUC of 0.85, distinct from the fold-wise macro ROC–AUC of 0.816 ± 0.083.The figure also presents one-vs-rest ROC curves, precision–recall curves, and a normalized confusion matrix for SD, PsP, and TP.Source: Figure A.1 · p. 30