AXONS-3: An XAI-Augmented Approach for Advancing Trust and Transparency in 3D Brain Tumor Segmentation

Early brain tumor detection remains a critical challenge in medicine due to its impact on patient outcomes. While magnetic resonance imaging (MRI) is a key tool, the challenges accumulated from the grayscale nature of MRI and high volume of data, coupled with human cognitive limitations and time pressures in radiology, create a potentially large margin of diagnostic uncertainty and error. Despite their performance, deep learning solutions face resistance in clinical adoption due to the lack of trust that roots from the opacity of such models. This research introduces the AXONS-3 workflow with the aim of bridging model outputs with the practical needs of clinicians by integrating interpretability and transparency into artificial intelligence (AI) systems for clinical decision-making. First, a 3D U-Net model is trained on the BraTS2020 dataset using T1Gd, T2, and FLAIR MRI sequences to segment brain tumors into sub-regions of NCR/NET, ED, and ET. Then, post-hoc visual Explainable AI (XAI) techniques, including gradient-based methods and uncertainty quantification, are augmented to the workflow to interpret the process of reaching the predicted segmentation. The proposed AXONS-3 workflow provides visually intuitive feedback and justifications to foster greater stakeholder comprehension and trust, contributing to the transparency of AIdriven systems needed for reliable adoption in clinical settings.