AI-based opportunistic risk model to predict pancreatic cancer risk using pre-diagnostic contrast-enhanced CT scans.

Abstract

e16470 Background: Pancreatic cancer (PC) remains a leading cause of cancer-related mortality. While surgical resection offers the best chance for survival, patients are often diagnosed late, at an inoperable stage. Consequently, identifying high-risk individuals for surveillance is critical for improved survival. While recent efforts have utilized electronic health records for early risk-stratification, the predictive value of CT scans within an opportunistic screening framework remains under-explored. We hypothesized that (a) radiomic features capturing structural changes, and (b) quantification of steatosis (fat) capturing obesity and metabolic dysfunction from across different pancreatic subregions (head, body, tail), may be able to reliably predict likelihood of PC risk, up to 3-months before disease manifestation. Methods: We curated a dataset on n = 190 studies from the University of Wisconsin Hospital containing longitudinal pre-diagnostic CT scans with IV contrast (67 patients who developed PC within a timeframe of 36-months, 123 healthy patients). Scans were reorientated, reshaped, HU’s clipped to [-200,300], and normalized. Pancreas and subregion segmentations were generated via PanSegNet. Radiomic features (intensity, laws, gradient, haralick, CoLIAGe, shape) as well as fat (ie. steatosis) features from different subregions (head, body, tail) were computed per scan. Highly correlated and low variant features were removed before further pruning via elastic net. Following cross-validation within a random forest model with 70% for training, 15% for validation, and 15% for testing, a binary classification was performed for forecasting malignancy in 3-12 months, 24-36 months, and 3-36 months. Results: A combination of 20 features including law textural features, perimeter, convexity, volume, and voxel-wise counts of fat within the pancreas and its subregions along with fat distribution discrepancies between different regions resulted in AUROCs of 76.2% ± 9.1% for 3-12 months, 77.9% ± 11.7% for 24-36 months, and 69.3% ± 9.1% for 3–36 months on the test set. These results suggest that there may be a relationship between pancreas dysfunction and the spatial distribution of pancreas fat and pancreas morphology. Conclusions: Radiomics and fat features from opportunistic CT scans may be complimentary to existing risk-stratification approaches in flagging patients at risk for PC. A large multi-institutional validation is warranted.