Detecting pancreatic tumors with artificial intelligence and machine learning

Introduction: Pancreatic cancers are tumors with poor prognosis due to late presentation of clinical symptoms and lack of screening tests. Artificial intelligence technologies are revolutionizing the field of huge data analysis. Faster Regional Convolutional Neural Networks (Faster R-CNN) stand out among image recognition algorithms for object detection with high accuracy, sensitivity, and specificity scores with fast results. Aim: In our study, we aimed to establish early diagnosis of pancreatic tumors by training an artificial intelligence model with computed tomography images of patients with pancreatic tumors recorded in our center; to automate the process and reduce the workload of experts in regions where access to medical doctors is difficult; and to reduce the burden on the current healthcare system, Materials and Methods: Total of 710 patients, including 350 patients whose pathologies were confirmed with pancreatic tumors and 360 patients with healthy pancreas images, who were admitted to the Department of General Surgery, Cerrahpaşa Faculty of Medicine, Istanbul University Cerrahpaşa, between 2016 and 2024, were included in the study. A supervised learning artificial intelligence model was created using labeled Computed Tomography images with the Faster R-CNN deep learning tool. Results: Our model was able to localize and diagnose tumors in 48 of the 50 patients with pancreatic tumors confirmed by imaging and pathology in the test set and categorized 197 of the 200 healthy pancreas reports as healthy. The accuracy of the model was measured as 98.4%, the sensitivity as 98%, and the specificity as 98.5%. The positive likelihood ratio was 65.33, and the negative likelihood ratio was 0.02. The F1 score of the model was calculated as 95.04%. In the test set, the classification CNN predicting pancreatic tumors was found to be positively correlated with positive OO (P = 0.04). The area under the curve was calculated as 0.998. Conclusion: Our model showed high accuracy, sensitivity, and specificity in detecting pancreatic tumors in Computed Tomography image scans. High Dice Score, F1 score, and EAA metrics are indicators of the undoubted usefulness of the model we developed.