PII-009 - DECIPHERING TYPE 1 DIABETES BY INTEGRATING ARTIFICIAL INTELLIGENCE WITH PANCREATIC WHOLE SLIDE IMAGING

O. Elashkar¹, N. Waddington¹, J. Penaloza Aponte², S. Kang³, H. Ju⁴, M. Lotfollahi⁴, D. Lamb⁵, M. Campbell-Thompson², S. Kim¹; ¹Department of Pharmaceutics, College of Pharmacy, University of Florida, ²Department of Pathology, Immunology, and Laboratory Medicine, Diabetes Institute, College of Medicine, University of Florida, ³University of Florida College of Pharmacy Department of Pharmaceutics, ⁴NVIDIA, ⁵University of Florida.

Background: Type 1 diabetes (T1D) is characterized by the destruction of pancreatic insulin-secreting β cells of the islets of Langerhans by autoimmune cells, leading to hyperglycemia. Islet autoantibodies are used as biomarkers for T1D risk and diagnosis. Despite the fact that the increase of autoantibodies increases the risk of diabetes, there are no conclusive diagnostic criteria for this association. Whole slide imaging (WSI) has revolutionized the field of pathology by being able to capture high-resolution, full-size organ images. We aim to correlate clinical and biochemical patient characteristics with α and β cell heterogeneity, providing valuable insights for diagnostic decision-making through artificial intelligence (AI).

Methods: 853 WSI were obtained from nPOD for 3 groups: non-diabetic autoantibody-negative controls (ND, n=16), autoantibody-positive non-diabetic (AAb+, n=16) and type 1 diabetic (T1D, n=35). Annotation was created over multiple rounds for total section, endocrine, and CD3+ areas. First, computer vision methods offered by OpenCV and Scikit-image have been leveraged. Then, manual refinement was done in QuPath with the aid of MONAI label and segment-any-thing (SAM) tools. HoVer-Net, a convolutional neural network capable of instance segmentation and classification, was constructed using MONAI and PyTorch as the main segmentation model. The resultant model has been used to infer and segment the rest of unlabeled images. The resultant imaging features—i.e., glucagon, insulin, and CD3 cell area—were analyzed with 4 tree-based machine learning methods to infer the extracted imaging features from clinical and biochemical characteristics. After cross-validation and tuning, the best approach has been selected based on RSME metric.

Results: The segmentation results have achieved a dice score accuracy >80%. A multi-output XGboost model has been selected to explain heterogeneity and infer the image pancreatic composition with RMSE < 0.23.

Conclusion: Our work shows the capability of AI in the emerging area of digital pathology and holds promising preliminary results for better diabetes understanding, targeted drug optimization, and monitoring for regenerative therapeutics. The output of this work would be useful for the clinical prediction of pancreatic status targeting more precision medicine.