MCMA-Net++: Topology-Aware and Graph-Driven Glioma Segmentation in 3D MRI

Jihan Alameddine, Céline THOMARAT, Remy Guillevin, Christine Fernandez-Maloigne, Carole Guillevin

Keywords: Glioma segmentation, Topology-aware learning, Graph neural networks, Hybrid CNN-Transformer, Medical image analysis

Abstract:

Glioma segmentation in 3D MRI remains challenging due to tumor heterogeneity, intensity
variability, and hierarchical anatomical structure. We propose MCMA-Net++, which
synergistically combines hybrid CNN-Transformer encoding, graph-based spatial reasoning
with anatomical priors, and a practical multi-component topology-aware refinement loss
tailored for nested tumor subregions. Our framework integrates: (1) Topology-Aware Refinement
Loss (TAR-Loss), enforcing consistency across nested subregions (ET, TC, WT),
and (2) Multi-Scale Anatomical Graph Reasoning (MSAGR), modeling spatial dependencies
through learnable graphs with anatomical priors. Combined with dual-stream CNNSwin
Transformer encoding and Multi-Class Multi-Attention, MCMA-Net++ achieves Dice
scores of 0.970±0.003 (WT), 0.943±0.005 (TC), 0.926±0.008 (ET), reducing HD95 from
5.48 mm to 3.21 mm compared to MCMA-Net. Graph reasoning contributes +1.3% Dice
for ET and TAR-Loss reduces topology violations by 41%. These results demonstrate the
effectiveness of combining topology-guided refinement and anatomical graph reasoning for
clinical-grade glioma segmentation.

Primary Subject Area: Segmentation

Secondary Subject Area: Geometric Deep Learning

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 97