Abstract A037: Predicting pancreatic cancer using artificial intelligence analysis of pancreatic subregions using computed tomography images

Abstract Study background: Early detection of pancreatic ductal adenocarcinoma (PDAC) can elevate the current ~10% five-years survival rate of PDAC up to 50%. Accurate stratification of high-risk individuals for PDAC can improve early detection as follow-up screening may assist diagnosis at an early stage. Studies show that the pancreas adopts changes prior to or during the development of cancer due to the underlying biological variations. This study aimed to examine the precancerous changes that occurred within and across pancreatic subregions to help stratify individuals at high risk of developing PDAC. Dataset: In a multi-institute retrospective study, 108 contrast-enhanced CT abdominal scans were collected, consisting of 36 diagnostic scans with established PDAC and observable tumor, 36 pre-diagnostic scans of the same subjects as in the diagnostic group but were obtained up to 3 years before PDAC diagnosis and were deemed ‘normal’ by radiologists, and 36 healthy scans reported with no PDAC signs. Trained radiologists outlined 3 subregions (head, body, tail) in all scans. Also, the subregions in pre-diagnostic scans were classified into high-risk (with cancer underdevelopment) and low-risk (no cancer development) groups by exploring the tumor signs in their corresponding subregions in the diagnostic scans. Experiments and results: Radiomic analysis was performed on all 324 subregions by extracting and analyzing hundreds of morphological and textural features. In a pairwise feature analysis (i.e. between corresponding subregions), the texture of the high-risk subregions in pre-diagnostic scans was found significantly unique and statistically different than that of the low-risk subregions, supporting the study hypothesis. Such textural features are usually too minute and remain obscured when the pancreas is observed as a single structure. The analysis showed that AI can efficiently identify and quantify such predictors. A Naïve Bayes model was then trained using the same data to automatically predict PDAC using the textural features of the pancreatic subregions. In four-fold cross-validation, the model obtained prediction accuracy by correctly classifying pre-diagnostic and healthy CT scans by 88.2% on average, with sensitivity (true positive rate) and specificity (true negative rate) reaching 82.5% and 94.0%, respectively. The results of this preliminary study are promising and encouraging to further validate the model on a larger dataset. The model showed improved results over those produced in our recent study [1] in which the pancreas as a single structure was examined. The prediction based on the proposed model can potentially assist clinicians to undertake specialized screening, diagnosis, and treatment planning accordingly as the tumor structure, symptoms, and drug response for each pancreatic subregion differs a lot. 1. Qureshi et. al, Predicting pancreatic ductal adenocarcinoma using artificial intelligence analysis of pre-diagnostic computed tomography images. Cancer Biomarkers, 33(2), pp.211-217, 2022. Citation Format: Sehrish Javed, Touseef Ahmad Qureshi, Srinivas Gaddam, Ashley Wachsman, Linda Azab, Vahid Asadpour, Wansu Chen, Bechien Wu, Yibin Xie, Stephen Pandol, Debiao Li. Predicting pancreatic cancer using artificial intelligence analysis of pancreatic subregions using computed tomography images [abstract]. In: Proceedings of the AACR Special Conference on Pancreatic Cancer; 2022 Sep 13-16; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2022;82(22 Suppl):Abstract nr A037.