Combating false negatives in pancreatic cancer: A deep learning approach for aiding fine needle aspiration via accurate subregion identification

Abstract

Accurate pancreatic cancer diagnosis based upon Computed Tomography (CT) guided Fine needle aspiration (FNA), crucially depends upon segmentation and classification of cancerous subregions (head, body, and tail). This experiment proposes an Artificial Intelligence (AI) driven deep learning framework that integrates novel Pancreatic U-Network (PanUNet) for pancreatic subregion segmentation and Residual Network (ResNet50) with Squeeze-and-Excitation (SE) blocks for classification. The AI model was trained with 2895 slices and refined through data augmentation techniques. The segmentation performance was assessed with dice similarity coefficient, intersection over union, sensitivity, and specificity, whereas F1-score, precision, recall and root mean squared error were used to evaluate classification performance. The model achieved 96.46 % dice similarity coefficient and 98.96 % classification accuracy. The experimental results demonstrate enhanced feature extraction and improved classification accuracy with SE block integration. Compared to individual optimizers, a Mixed-Adaptive moment estimation- Root mean square propagation- Stochastic gradient descent (MARS) optimization technique aided in achieving superior performance in the proposed framework. An extensive comparative analysis of the proposed model against established methods showcasing significant improvements in segmentation and classification proves its potential for clinical applicability.