RSNT-CHNN: A Novel Deep Learning Framework for Pancreatic Cancer Detection and Classification on CT Images Using Transfer Learning

Received: 26 March 2026 | Revised: 3 May 2026, 8 May 2026, and 11 May 2026 | Accepted: 15 May 2026 | Online: 18 May 2026

Corresponding author: E. Dhishya

Abstract

Pancreatic cancer remains one of the most dangerous cancers worldwide due to its aggressive progression, late-stage diagnosis, and limited detectability of early pathological changes in Computed Tomography (CT) imaging. Therefore, it is essential to provide early and accurate diagnoses with CT scans to enable timely medical interventions and improve survival rates. This study presents a novel stage-by-stage deep learning framework for the detection and classification of pancreatic cancer based on CT images, using a ResNet50-based Convolutional Hopfield Neural Network (RSNT-CHNN) to increase both diagnostic precision and reliability. The first step involves preprocessing to enhance image quality and eliminate any distortion present in CT images. Noise removal is achieved through a Wiener filter, which preserves structural data while removing noise in an efficient manner. The next step applies Contrast-Limited Adaptive Histogram Equalization (CLAHE) to increase contrast within the pancreas and highlight subtle differences between normal and abnormal tissues. In the segmentation step, a TransUNet-based segmentation model extracts the pancreas region from the CT images, effectively capturing the complex spatial features of the images and providing a more precise localization of the areas where possible cancerous regions exist. After segmentation, a deep feature extraction step is conducted via the ResNet50 architecture, which extracts robust hierarchical features from the CT images of the pancreas. These deep features are then input into the CHNN classifier, which uses convolutional learning and associative memory to accurately classify pancreatic cancer stages. Experimental results show that the proposed RSNT-CHNN framework reaches a classification accuracy of 98.10%, surpassing other state-of-the-art deep learning models. The results also demonstrate that combining advanced preprocessing techniques, a transformer-based segmentation model, deep residual feature extraction, and a Hopfield-based classification model can result in significant enhancements in the detection of pancreatic cancer. The proposed framework can be used as a highly effective computer-aided diagnostic tool to assist radiologists in the early detection and stage-wise classification of pancreatic cancer based on CT imaging.