Dual-path Learning via Optimal Transport Fusion for Precise Brain Tumour Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumour segmentation in multimodal multiparametric magnetic resonance imaging (MRI) is crucial for clinical decisions.
• Current U-Net architectures face challenges in capturing global context and delineating complex tumour boundaries, limiting clinical utility.

Purpose of the Study:

• To develop a novel dual-path 3D U-Net framework integrating Optimal Transport (OT) theory for enhanced brain tumour segmentation.
• To improve the alignment of global and local features for more accurate delineation of tumours with irregular or diffuse boundaries.

Main Methods:

• A dual-path 3D U-Net architecture with a Global Context Path (Pyramid Pooling) and a Local Detail Path (multi-scale convolutions).
• An Optimal Transport Fusion module for principled alignment and merging of features from both paths.
• An edge-aware loss function using 3D Sobel operators to refine segmentation mask boundaries.

Main Results:

• The proposed model demonstrated superior performance compared to standard U-Net and other state-of-the-art models on the BraTS 2023 glioma dataset.
• Optimal Transport-based feature fusion significantly improved the alignment of global and local features.
• The edge-aware loss function enhanced boundary precision in segmentation masks.

Conclusions:

• The novel dual-path 3D U-Net with Optimal Transport fusion offers a principled and effective approach to brain tumour segmentation.
• This framework addresses limitations of existing methods in capturing context and handling complex tumour morphologies.
• The findings highlight the potential of OT in advancing medical image analysis and improving clinical outcomes.