A Lightweight Attention based Deep Learning Framework for Pancreatic Cancer Detection in CT Images

Pancreatic Cancer is one of the deadliest cancers, with a five-year survival rate of about 11% because it is often diagnosed late. Current deep learning methods for detecting Pancreatic Cancer from Computed Tomography (CT) images frequently struggle with overfitting, high computational demands, and poor multi-scale feature representation. This paper presents a Lightweight Convolutional Neural Network (LWCNN) framework based on the Electrophorus Chain Hunt Algorithm (ECHA) for automated PDAC detection. In this study, firstly, an efficient segmentation of the pancreas and multi-level feature extraction from the image using a BAM-based U-Net architecture are performed. Subsequently, Modified Local Ternary Pattern (MLTP), Statistical Texture Pattern (STP), and LayerCAM features are integrated to construct a comprehensive feature representation. The fused features are classified by a light-weight CNN with the help of ECHA. The proposed framework is evaluated on the publicly available Medical Segmentation Decathlon (MSD) dataset, which comprises 420 three-dimensional CT volumes. The experimental results showed 95.80% accuracy, 96.39% sensitivity, and 95.20% specificity. Additionally, 10-fold cross-validation resulted in an accuracy of 94.05%, demonstrating solid robustness and generalization. Comparative and ablation studies confirm that integrating BAM-based segmentation, hybrid feature fusion, and ECHA optimization improves pancreatic cancer detection performance while keeping computational efficiency.