Enhancing Pancreatic Cancer Diagnosis with Lesion-Aware Graph Reasoning and Deep Feature Integration

Detecting pancreatic cancer early enough is still a tuff challenge because the tumors are usually small, have low contrast, and can hardly be distinguished from the surrounding anatomical structures in CT images. In this paper, the authors present a lesion-aware graph-augmented deep learning system that integrates physics-guided preprocessing, hybrid U-Net segmentation, radiomic feature extraction, MobileViT-based representation learning, graph reasoning, feature fusion, and probability calibration for pancreatic cancer detection. The system was tested with 1,418 high-resolution CT slices from the Pancreatic CT Images dataset. The proposed method has demonstrated a 94.2% accuracy at the slice level, 93.6% macro-F1 score, AUROC of 0.972, and AUPRC of 0.969. At the patient level, the results were further enhanced, reaching 96.0% accuracy, 95.7% macro-F1, 0.986 AUROC, and 0.983 AUPRC. Temperature scaling has decreased the Expected Calibration Error from 0.031 to 0.009, This way making the prediction more reliable. The segmentation network was scored by Dice at 0.82 0.09 and HD95 at 7.3 mm. Ablations have established that lesion-aware attention, graph learning, and radiomic features are very important to the performance. The system was also tested for strength when it comes to imaging artifacts and cross-site variations, and it performed outstandingly with AUROC values consistently higher than 0.94. Besides, the system achieved efficient deployment with an average inference time of 28 ms per slice and 1.48 s per study. Such results indicate that the combination of lesion-aware attention, graph-based contextual learning, and calibrated deep representations can give an accurate, reliable, and computationally efficient solution to the problem of pancreatic cancer detection from CT images. Keywords: Pancreatic cancer; Temperature scaling; Lightweight graph head; Lesion-aware feature fusion; post-hoc probability calibration.