S227 Accuracy of Artificial Intelligence in Imaging Modalities to Detect Pancreatic Neoplasms: A Systematic Review and Meta-Analysis

Introduction: Early identification and detection of pancreatic cancer significantly increases the median overall survival rate of patients. However, this is challenging since routinely utilized genetic tests are costly and time intensive. Artificial intelligence (AI) is an emerging technology being used for interpretation of medical imaging. AI can aid in early and quick diagnosis of pancreatic cancer through analysis of medical images like CECT and ultrasound which can contribute as a capable screening tool for the requirement of further testing. This study analyzes the accuracy of different models of AI in the diagnosis of pancreatic neoplasms. Methods: The study follows the PRISMA guidelines and medical databases were searched by creating a Boolean search string to retrieve and select articles which assessed the accuracy of machine learning, deep learning and neural network models in diagnosis of pancreatic neoplasms. The meta, metadata and mada packages were utilized to evaluate the measures of diagnostic accuracy tests ie; Sensitivity, Specificity and diagnostic odds ratio (DOR). To evaluate sensitivity, the number of events was considered to be TP and the number of samples to be TP + FN. Similarly, pooled specificity was estimated by considering TN as events and TN and FP as the number of samples. The random intercept logistic regression model for sensitivity and specificity and the Mantel-Haenszel method or Inverse variance method was used to determine diagnostic odds ratio. The heterogeneity of the studies was assessed using the I2 test. Results: The study involved 20 different testing and validation models which assessed the diagnostic accuracy of machine learning and deep learning models. The pooled sensitivity was estimated by the random effects model to be 91% (88%-94%; 95% CI; P < 0.05) and the specificity to be 92% (87%-95%; 95% CI; P < 0.05). The diagnostic odds ratio was estimated to be 4.05 (2.49-6.59; 95% CI; P < 0.05). Conclusion: AI through machine learning and deep learning techniques, shows considerably high Sensitivity (91%) and Specificity (92%) in the detection of pancreatic carcinoma. Outcomes derived by AI had high correlation with histopathological data, establishing the role of AI in diagnosis of pancreatic cancer. In order to implement AI technology for diagnosis of Pancreatic cancer, additional multicenter studies must be performed and DDW 2025 involvement of multiple disciplines to implement this technology is essential.