Deep learning-based CT-scan image classification for accurate detection of pancreatic cancer: A Comparative Study of Different Pre-Trained Models

Pancreatic cancer remains one of the deadiest forms of cancer worldwide. The main challenge in pancreatic cancer diagnosis is primarily attributed to the late stage at which it is typically diagnosed. Computed tomography (CT) has been used to provide a concise visualization of the pancreas, however, due to late pancreatic tumor diagnosis, this imaging technique is yet to be highly effective in detecting this malignancy at an early stage. Therefore, this research aims to apply pre-trained deep learning models in classifying CT scan images for the early detection of pancreatic cancer. Residual Network 50 (ResNet50), MSHA, and EfficientNet are the algorithms applied to analyze CT scans and identify cancerous growths in the pancreas. The results show high classification precision, with the highest achieved precision score of $\mathbf{1 0 0. 0 0 \%}$ using MSHA. However, all the applied models had a high performance in detecting pancreatic cancer as they all had scores not less than $\mathbf{9 9 \%}$ in all performance metrics used. These findings emphasize its proficiency in precisely identifying and categorizing pancreatic cancer. This program aims to alleviate the tremendous workload of healthcare systems caused by the large number of medical images analyzed by medical professionals. It specifically can assist radiologists and other specialists in detecting pancreatic tumors by offering a faster and more accurate technique.