Bridging Accuracy and Interpretability in Deep Learning for Pancreatic Tumor Segmentation

Pancreatic cancer is also one of the most lethal ones as it is often diagnosed at its last stage, or it is hard to detect small tumors during medical examination. Deep learning has demonstrated potential in the automatic segmentation of pancreatic tumors using computed tomography (CT) images; nontransparent decision-making of these models is the barrier to clinical use. The paper introduces a basic CNN network to pancreatic tumor segmentation on the U-Net architecture based on Dice loss to overcome class imbalance. To tackle novelty, the suggested system uniquely combines Grad-CAM explainability with anatomy-constrained post-processing in a 2D U-Net baseline. This is a method that has not been used in previous 2D pancreatic segmentation baselines. It offers both quantitative tumor overlap metrics and qualitative attention analysis. According to the findings of the experiments, although the model reveals the global anatomical structures, the localization of the small tumor areas is not yet accurate. The visualizations of grad-cam indicate that the early-stage models pay attention to the general abdominal areas, but do not concentrate on the tumor areas, which proves that pancreas-oriented and context-inspired learning is necessary.