MALCNet: modality-aware latent completion network for brain tumor segmentation with incomplete MRI modalities

Le résumé fourni par la source

Accurate brain tumor segmentation from multi-modal magnetic resonance imaging (MRI) relies on complementary information from T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and FLAIR sequences. In clinical practice and retrospective studies, one or more MRI modalities are often missing due to heterogeneous acquisition protocols, patient-related constraints, or incomplete historical records. Such modality incompleteness introduces distribution shifts that substantially degrade the performance of segmentation models trained on complete multi-modal inputs. We propose MALCNet, a modality-aware latent completion network for robust brain tumor segmentation under arbitrary missing MRI modalities. MALCNet completes missing information in the latent space through three key components: a shared-weight encoder with an adaptive modality-specific correction unit to extract anatomy-consistent yet modality-sensitive hierarchical features; a modality-aware imputation transformer that performs bottleneck-level completion, initializing missing modalities with learnable semantic priors and refining them via cross-modal attention; and an attention-guided multi-scale propagation module that transfers completed high-level semantics to shallow feature maps to recover fine-grained spatial details. During training, a dual-branch consistency optimization strategy aligns predictions and features between complete- and incomplete-modality inputs. MALCNet is evaluated separately on two BraTS releases (BraTS2019 and BraTS2020) under all 15 non-empty modality configurations. It achieves average Dice similarity coefficients of 86.68%, 76.73%, and 63.60% for whole tumor, tumor core, and enhancing tumor, respectively, on BraTS2019; the corresponding values on BraTS2020 are 87.86%, 79.59%, and 64.10%. Subject-level paired comparisons against six incomplete-modality methods show statistically significant differences for all tumor regions on both releases (all P -values ≤ 0.002). MALCNet provides a robust approach for brain tumor segmentation with incomplete MRI modalities. By combining modality-aware latent completion with multi-scale feature propagation and dual-branch consistency optimization, the model effectively preserves both anatomical coherence and modality-specific tumor cues, achieving reliable segmentation across diverse missing-modality configurations.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.