Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive cancer marked by substantial intra-tumor heterogeneity, complicating early detection and treatment efforts. This study introduces a novel, modified U-Net model designed for efficient and cost-effective tumor segmentation and analysis of intra-tumor heterogeneity in PDAC. The model achieves an impressive segmentation accuracy of 96.78%, with fivefold cross-validation yielding a mean accuracy of 96.2% and a variance of 0.052. Radiomic features—including intensity, texture, and shape-based characteristics—were extracted from the segmented regions. Otsu-based clustering was applied to visualize distinct heterogeneous regions within the tumor, emphasizing intensity and texture pixels to highlight variability. The results were clearly visualized, offering enhanced insights into the clustering outcomes and intra-tumor heterogeneity. These findings provide valuable information for accurate diagnosis and personalized treatment planning, showcasing the potential of this AI-driven model to significantly improve medical imaging in a practical and efficient manner.