TTG-U-Net: An Interpretable and Efficient Framework for Multi-Modal Brain Tumor Segmentation Enabling Clinically-Aligned Decision Support.
Authors
Abstract
Accurate, automated analysis of multi-modal MRI is foundational to neuro-oncology informatics. However, the integration of Artificial Intelligence (AI) into clinical workflows is critically hampered by the "black box" nature of many deep learning models, which erodes clinical trust and complicates validation. This challenge is compounded by the difficulty of developing integrative approaches that can transparently and efficiently fuse heterogeneous information from multiple MRI sequences. To address this critical gap in Explainable AI (XAI) for healthcare, we propose TTG-U-Net, a novel segmentation framework designed to bridge performance with clinical interpretability and efficiency. Our framework provides an integrative solution through three synergistic components: (1) a cross-modal Transformer that explicitly models inter-modality dependencies, yielding attention maps that serve as a transparent visual audit trail for the model's reasoning, directly addressing the need for XAI; (2) a dynamic low-rank tensor decomposition that adaptively regularizes the model and reduces its computational footprint, facilitating deployment in standard hospital information systems; and (3) a modality-adaptive gating mechanism that learns a transparent information routing policy, mimicking established radiological principles. Validated on the BraTS 2021 benchmark, a single TTG-U-Net achieves state-of-the-art performance (Dice: WT 91.7%, TC 88.8%, ET 84.5%), competitive with computationally-intensive ensembles. The dynamic low-rank design reduces the parameter count by approximately 41%, enhancing efficiency for practical deployment. Crucially, interpretability studies confirm the model's learned fusion strategy aligns with clinical knowledge, bolstering its trustworthiness. TTG-U-Net offers a compelling framework that moves beyond pure segmentation accuracy, providing a robust, efficient, and trustworthy tool poised for meaningful integration into clinical decision support systems and informatics workflows, demonstrating a viable path for responsible AI in healthcare decision-making.