TTG-U-Net: An Interpretable and Efficient Framework for Multi-Modal Brain Tumor Segmentation Enabling...

Area of Science:

• Neuro-oncology informatics
• Artificial Intelligence (AI) in healthcare
• Medical image analysis

Background:

• Accurate, automated analysis of multi-modal MRI is crucial for neuro-oncology.
• Clinical integration of AI is hindered by the 'black box' nature of deep learning models, impacting trust and validation.
• Developing transparent and efficient methods to fuse heterogeneous MRI data remains a challenge.

Purpose of the Study:

• To propose TTG-U-Net, a novel segmentation framework addressing the need for Explainable AI (XAI) in healthcare.
• To bridge segmentation performance with clinical interpretability and computational efficiency.
• To facilitate the integration of AI into clinical workflows for neuro-oncology.

Main Methods:

• A cross-modal Transformer to model inter-modality dependencies and generate attention maps for transparency.
• Dynamic low-rank tensor decomposition for adaptive regularization and reduced computational footprint.
• A modality-adaptive gating mechanism for transparent information routing, aligning with radiological principles.

Main Results:

• TTG-U-Net achieved state-of-the-art performance on the BraTS 2021 benchmark (Dice: WT 91.7%, TC 88.8%, ET 84.5%), competitive with ensembles.
• The dynamic low-rank design reduced the parameter count by approximately 41%, enhancing efficiency.
• Interpretability studies confirmed the model's fusion strategy aligns with clinical knowledge, increasing trustworthiness.

Conclusions:

• TTG-U-Net provides a robust, efficient, and trustworthy framework for multi-modal MRI segmentation in neuro-oncology.
• The model demonstrates a viable path for responsible AI integration into clinical decision support systems.
• Explainable AI in TTG-U-Net enhances clinical trust and facilitates validation for neuro-oncology informatics.