LASSNet: Level-Aware Availability-Conditioned Spatial-Semantic Fusion for Brain Tumor Segmentation with Missing MRI Modalities

• Missing sequences can arise from acquisition cost, protocol variation, artifacts, contraindications, or patient motion, motivating segmentation from any non-empty modality subset.Source: Introduction · p. 1
• Prior missing-modality work includes reconstruction, knowledge transfer, and direct fusion, while the introduction positions LASSNet around availability- and level-conditioned fusion.Source: Introduction · p. 1
• The introduction identifies representation-level variation as an open fusion question: lateral features preserve detail, while bottleneck features encode semantic and inter-modality context.It links missingness to changing feature magnitude and channel reliability in lateral paths, and to the value of relation modeling at compressed bottlenecks.Source: Introduction · p. 1
• LASSNet assigns complementary roles to HACF for count-normalized, mask-conditioned lateral fusion and TriRSF for available-modality relations and spatial context across bottleneck resolutions.A shared decoder starts from TriRSF semantics and progressively incorporates HACF detail, without synthesizing missing sequences.Source: Introduction · p. 1
• 76.7% and 83.2% mean WT/TC/ET Dice scores are reported on BraTS2019 and BraTS2023, respectively, across all 15 non-empty modality configurations.Source: Introduction · p. 2

• Reconstruction and representation-transfer methods address missing modalities through synthesized evidence or constraints on incomplete representations, with generation, teacher, or alignment objectives.Source: Related Work · p. 2
• Direct latent-space methods aggregate observed features using statistics, gated or region-aware fusion, self-attention, Transformer interactions, or state-space approaches.The passage notes that these methods commonly apply one fusion family at a selected site or repeatedly across stages.Source: Related Work · p. 2
• Prior level-aware work distinguishes network stages, including Hierarchical In-Out Fusion with shallow self-attention and deepest-layer spatial–frequency fusion.Source: Related Work · p. 2
• LASSNet’s distinction is availability-conditioned level-specific organization: HACF fuses and locally refines lateral features, while TriRSF models modality relations and spatial context across three bottleneck scales.The paths are coupled through coarse-to-fine decoding.Source: Related Work · p. 2

Problem Formulation and Framework Overview

• LASSNet handles any non-empty modality subset with parallel availability-conditioned lateral and bottleneck paths, then decodes from TriRSF semantics while injecting HACF detail without reconstructing missing sequences.TriRSF does not take HACF as its input; the paths meet through decoding.Source: Problem Formulation and Framework… · p. 2, Figure 1 · p. 3

Hierarchical Availability-Conditioned Fusion

• HACF averages observed features with count normalization, applies availability-conditioned channel modulation at the first three levels, and refines the fourth level with local attention and a spatial feed-forward network.The four lateral features retain their native decoder alignment; cross-scale semantic interaction is left to TriRSF.Source: Hierarchical Availability-Conditioned… · pp. 3–4

TriRSF: Tri-Scale Relational–Spatial Fusion

• TriRSF builds three bottleneck scales, scores available-modality descriptors, models descriptor relations, consolidates local spatial features across scales, then applies global attention over spatial tokens.The availability embedding records the contributing subset after modality aggregation.Source: TriRSF: Tri-Scale Relational–Spatial… · pp. 4–5

Learning Objective

• Training supervises modality-specific auxiliary branches on complete scans, adds ET-aware Tversky and gated residual-error losses, and periodically uses complete-mask predictions as consistency targets; only the fused prediction is retained at inference.The sampled availability mask conditions the fused path, while every modality-specific auxiliary branch receives supervision.Source: Learning Objective · p. 6

Implementation Details

• Training uses patient-disjoint splits of 234/33/68 BraTS2019 cases and 875/125/251 BraTS2023 volumes, with one of 15 non-empty availability masks sampled per training case.The implementation uses random 128^3 crops and batch size one.Source: Experiments · p. 6

Evaluation Protocol

• BraTS2019 uses the terminal checkpoint after 1000 epochs, whereas BraTS2023 continues training through epoch 1680 and selects the best validation checkpoint.Testing covers all 15 masks using 128^3 sliding windows at 50% overlap; Dice is averaged over subjects and then masks for WT, TC, and ET.Source: Experiments · p. 7

Comparison with Existing Methods

• On BraTS2019, LASSNet's overall Dice is 76.7%, exceeding mmFormer by 0.8 points; its ET, TC, and WT leads are 1.4, 0.6, and 0.5 points.LASSNet reports 61.9% ET, 80.4% TC, and 87.9% WT Dice.Source: Experiments · p. 7
• On BraTS2023, LASSNet reaches 83.2% overall, leading IM-Fuse by 0.3 points on WT, tying on TC, and trailing by 0.7 points on ET and 0.1 points overall.The configuration-wise pattern includes 91.9% WT Dice with FLAIR alone and ET Dice rising from 65.1% to 83.6% when T1ce is added to FLAIR–T1–T2.Source: Experiments · p. 7

Ablation Study and Visualization

• Under the matched 1000-epoch BraTS2019 schedule, Full reaches 76.72% Mean Dice versus Base at 73.84%, while neither specialized path alone surpasses Base.The joint configuration gains 2.88 points, with a 4.87-point ET gain; repeated runs are needed to quantify stability.Source: Experiments · p. 7, Table 3 · p. 7

Discussion and Conclusion

• LASSNet assigns lateral features to spatial detail and bottleneck features to available-modality relations and semantic context, rather than using one fusion operator at every level.The paths reconnect through a shared coarse-to-fine decoder without reconstructing missing inputs.Source: Discussion and Conclusion · p. 7
• Across all 15 non-empty configurations, LASSNet obtains mean Dice scores of 76.7% on BraTS2019 and 83.2% on BraTS2023, supporting availability-conditioned level specialization for incomplete multimodal segmentation.Source: Discussion and Conclusion · p. 7
• Under the matched 1000-epoch schedule, Full improves Mean Dice over Base by 2.88 points, with the largest gain on ET, supporting joint use of both fusion paths.Source: Discussion and Conclusion · p. 7
• Future work will examine repeated-run stability, naturally incomplete multi-center cohorts, uncertainty-aware prediction, and computational efficiency.Source: Discussion and Conclusion · p. 7