DEEP LEARNING CHARACTERIZATION OF MICROSCOPIC INTRADUCTAL PRECURSOR LESIONS AND ASSOCIATED PARENCHYMAL EFECTS IN INDIVIDUALS AT HIGH RISK FOR PANCREATIC CANCER

BACKGROUND: Detecting early PDAC remains a challenge due to the inability of current diagnostic tools to visualize small cancers and microscopic PDAC precursors (PanINs). PanINs are associated with changes in the adjacent pancreatic parenchyma, characterized by acinar atrophy, fibrosis, and fat(lobulocentric atrophy or LCA). Defining the histologic changes associated with PanINs may form the basis for understanding the subtle changes PanINs produce on EUS and allow for their clinical detection. AIM: We compared the tissue composition and PanIN burden of resected pancreata of high-risk individuals (HRI) with a familial/genetic predisposition for PDAC with non-high-risk controls. METHODS: HRI enrolled in prospective Cancer of the Pancreas Screening (CAPS) studies underwent periodic MRI and EUS. We compared scanned H&E-stained histological sections of resected pancreata from 34 HRI patients who had surgery for suspected neoplasms and 62 controls matched by age group, gender, smoking, and BMI group who had surgery for benign pancreatic conditions. Control specimens were chosen from a prospectively collected surgical pathology database. We used a novel deep learning model (CODA) trained on >10,000 examples of 9 pancreatic microstructures to segment and quantify the tissue composition in 1400 scanned images of XY slides for the non-malignant parenchyma. We calculated the proportions of 9 tissue components, including PanIN and LCA burden (PanIN+fat+collagen), in each patient’s resected pancreas. We created a spatial tissue heterogeneity score to quantify the range of variability of structural changes for each subject (Fig 1F-H). RESULTS: HRIs were comparable to controls with respect to age, gender, BMI, smoking, alcohol, and diabetes (Table 1). 33.4% of HRIs had PDAC/high-grade dysplasia (HGD) in their resection specimen (progressors). The proportion of the pancreatic tissue that was PanIN (Fig1A) and the proportion of ducts involved by PanIN (Fig 1B) was higher in HRIs than in controls (p<0.0001). The proportion of the pancreas that was PanIN-associated interlobular fat (p<0.01, Fig 1C) and collagen (p<0.0001, Fig 1D) was also greater in HRI. HRI and controls had comparable amounts of islets, nerves, and blood vessels. Interlobular fat content correlated with PanIN content in HRI (p=0.028) but not in controls (Fig 1E). PanIN and LCA burden strongly correlated with tissue heterogeneity in HRI (p=0.003) but not in controls (Fig1 F-H). PanIN and LCA burden were higher in HRI progressors than non-progressors (p=0.02, Fig 1I). CONCLUSION: HRIs have a greater proportion of PanIN, fat, fibrosis, and atrophy, resulting in pancreatic tissue heterogeneity, which may be detectable by diagnostic imaging. These quantitative metrics support ongoing deep learning approaches using radiomics (EUS, MRI) to improve early detection and prevention of PDAC in HRI.