77-PUB: Deep Learning–Powered Islet Quantification for Type 1 Diabetes

Introduction & Objective: Islets in patients with type 1 diabetes (T1D) display a high degree of heterogeneity in sizes, numbers, architecture, endocrine and immune cell type compositions, and capillary density, compared to nondiabetic controls (ND). Extensive whole slide imaging (WSI) data collected through the JDRF Network for Pancreatic Organ donors with Diabetes (nPOD) is an opportunity to analyze T1D progression. We aimed to enhance the efficiency of quantitative WSI analysis using deep learning.

Methods: Multiple computational approaches were explored, including OpenCV, Hover-Net, Segment Anything Model (SAM), and a pixel classifier. We leveraged HiPerGator, UF supercomputer, and QuPath, open-source software widely used for WSI analysis. We tested the developed workflow using 30 WSI data (two slides from the pancreas tail region for five subjects in three disease states: ND, autoantibody+, and T1D).

Results: Among the tested approaches, we selected two deep learning models, SAM and the pixel classifiers implemented in QuPath. SAM was employed to yield precise boundaries of the islets within the Region of Interest (ROI) boxes implemented by the stain classifier. The pixel classifier was further applied to segment the areas corresponding to insulin and glucagon within each islet. We automated counting CD3+ cells detected inside and within 20um outside around the perimeter of the islets. For the testing data, we acquired accurate quantitative measurements for the areas of insulin and glucagon inside islets and the count of CD3+ cells.

Conclusion: With the developed workflow, we enhanced the efficiency of quantifying features of the WSI data. It also assists in defining insulitis more accurately by counting the number of CD3+ cells. We intend to utilize our workflow to extract features from a large set of WSI obtained through the JDRF nPOD program. Analyzing extracted features from WSI will aid in understanding T1D progression.