Hybrid deep learning framework for pancreatic cancer detection and segmentation using graph attention CNN and swin transformer
Simon Tongbram, Nameirakpam Dhanachandra, Thelma Ngangom · ICT Express · 2025
This paper introduces PANDA-Net (Pancreas Detection and Adaptive Network Architecture), a hybrid deep learning framework for accurate pancreatic cancer detection and segmentation. PANDA-Net integrates Adaptive Contrast Enhancement with Residual Learning (ACERL) to recover fine structural details, Graph Attention-based CNN (GACNN) to preserve local and global spatial relationships, a modified Swin Transformer with Adaptive Token Pruning (Swin-ATP) to enhance classification efficiency, and Sand Cat Optimization (SCO) for dynamic hyperparameter tuning. The framework was evaluated on multimodal datasets, including CT, T1-weighted MRI, and T2-weighted MRI, using a five-fold cross-validation strategy. Results demonstrate that PANDA-Net achieves a Dice similarity coefficient of 99.55% and classification accuracy of 99.6%, outperforming baseline models such as U-Net, DeepLab V3, ResUNet, and Attention U-Net. The findings highlight PANDA-Net’s potential in addressing pancreas-specific challenges, offering improved segmentation precision, classification stability, and computational efficiency for clinical applications.