Visualizing UNet Decisions: An Explainable AI Perspective for Brain MRI Segmentation
D. Jeya Mala, Mainak Chattopadhyay, Parthiba Mukhopadhyay, Roopak Sinha · IEEE Access · 2025
In recent years, medical image analysis, particularly in neuroimaging, has experienced remarkable advancements, with magnetic resonance imaging that greatly helps in diagnosing complex neurologigal disorders, including brain tumors. However, accurately segmenting brain tumors from MRI scans remains a significant challenge, necessitating sophisticated computational techniques. This paper presents the outcomes of a research work focused on brain MRI segmentation utilizing the UNet architecture, which aims to harness its effectiveness in semantic segmentation tasks focusing on enhancing model interpretability using explainable AI methods.This paper dives deep into the intricacies of UNet’s adaptation to brain MRI segmentation, the dataset employed, and the methodology for model development, training, and validation. In addition to discussing the segmentation outcomes, we incorporate several explainable AI techniques that include Grad-CAM, Saliency Maps, Vanilla Gradient and Layer-wise Relevance Propagation (LRP) to obtain the necessary visualizations. Comprehensive analyses of the results highlight the clinical implications of these findings, addressing both the potential benefits and limitations of various XAI methods. By the application of XAI methods, this work obtained necessary visualizations, to recognize the internal working of the black box nature of the U-Net architecture. In addition to that, the comprehensive analysis of the results highlight the clinical implications of these findings, addressing both the potential benefits and limitations of different XAI methods in visualizing the model’s outcomes. Based on the analysis of the XAI evaluation metrics such as Fidelity, Unambiguity and Stability, the Vanilla Grad method stands out with its high unambiguity and consistent fidelity scores, making it a reliable choice for providing clear and interpretable explanations in complex scenarios. In addition, the findings suggest that while LRP offers solid stability, the combination of high fidelity and clarity from the Vanilla Grad model makes it the preferred method for enhancing the interpretability of AI systems in brain tumor segmentation. Overall, this research work represents a significant advancement in leveraging the trustworthiness of UNet architecture in accurate and efficient brain tumor segmentation by means of XAI methods, ultimately aiming to support clinicians in diagnosis and treatment planning while fostering a deeper understanding of the model’s decision-making processes.