Subject Area: COMPUTER SCIENCE
Pancreatic Cancer (PC) is among the deadliest forms of cancer, with a high mortality rate due to its often-late diagnosis. Early detection plays a critical role in improving survival rates, but this remains a challenge in medical imaging. This study presents an integrated deep learning-based approach for the automated tracking and detection of pancreatic tumors using a combination of Convolutional Neural Networks (CNN), as a feature extractor, ReSNet and Region Proposal Networks (RPN) for classification tumor localization in CT images. The proposed system aims to enhance diagnostic accuracy and assist radiologists in identifying early-stage tumors. The methodology used is Agile. The research methods involve data collection from Madely repository considering 201 patients of undergoing first-line surgery for Pancreatic Ductal Adenocarcinoma (PDA), across 25-75 years of age, and considering diverse stages of PC. Proposed CNNs with multi scale convolutional process was applied for to extract relevant features from CT scan images. The extracted features were applied to train ResNet as the classifier, while RPN is then used to detect potential regions of interest where tumors might be present. The combined model is evaluated on a pancreatic cancer dataset, where various metrics such as F1-score, precision, recall, and confusion matrices are used to assess its performance. The model achieved an F1-score of 0.97 with a confidence threshold of 0.395 and a recall of 1.0 at the same threshold. The confusion matrix indicates a 93% correct prediction for tumors, with the rest categorized under background and no-tumor classes. The mAP score, a key metric for object detection models, improved significantly from 0.1973 to 0.85331, demonstrating the model's robustness in identifying tumors. The results highlight the potential of deep learning models in medical diagnostics, particularly in addressing the complex challenge of early pancreatic tumor detection. The model was integrated as software and tested with real world PC CT image data. The results showed that our model was able to correctly classify the PC image and segment the infected region with 97% accuracy.