Objective
The overall objective of this project is to develop an Artificial Intelligence (AI)-based imaging analysis and informatics tool for better understanding pathogenic mechanisms in T1D for both endocrine and exocrine pancreas. A deep learning (DL) analysis will be conducted through optimizing the use of the substantial pancreas whole slide image (WSI) collection of the JDRF nPOD program amassed over 14 years at the University of Florida Diabetes Institute. As well, we will develop an AI 3D pancreas volume tool using existing and ongoing pancreas MRI scans from subjects enrolled in the NIH DP3 TrialNet study (1DP3-DK-101120-01, completed) and recently started longitudinal pancreas volume study (R01DK123329). Importantly, many of the subjects who were imaged for the DP3 observational study will be approached to participate in the longitudinal R01 study and indeed 8 subjects have already been enrolled though enrollment only started in July, 2021. The central hypothesis of this proposal is that architectural modifications of the pancreatic islets due to loss of β-cells and infiltration of the endocrine and exocrine compartment by CD3+T cells throughout the pancreas and of the exocrine 3D morphology can be characterized and heterogeneity can be quantified by AI-assisted imaging analysis. The outcome of these studies will further our understanding of the pathogenic mechanisms within the endocrine and exocrine pancreas that increase risk for T1D and provide regional morphometric parameters that inform interpretation of morphological features of pancreas from MRI imaging. Together these studies will support the use of pancreas MRI as a biomarker for T1D risk.
The specific aims are as follows:
Aim 1. Quantify the heterogeneity of the pancreatic islets and exocrine inflammation within and between individuals by developing an advanced pancreatic imaging analysis and informatics tool. Using DL-based segmentation algorithms, we will extract boundaries of the endocrine cells within pancreatic islets and single cells scattered in the exocrine region to provide endocrine single-cell counts and density and nearest neighbor distances. We will also quantify single CD3+T cells at islets (contouring within 20 µm and successive 20 µm intervals) and within the exocrine regions to quantify islet morphometry and insulitis status and exocrine infiltrates. We will examine these variables in normal controls, pre-T1D donors (single and multiple AAb+ nondiabetic), and patients with recent (≤1 year) and chronic durations of T1D using statistical analysis.
Aim 2. Quantify the longitudinal changes in pancreatic volume within and between individuals using AI algorithms and statistical modeling analysis. To estimate pancreatic volume, we will perform image segmentation on MRI images to determine the pancreas outline shown in each 2D slice using DL algorithms and construct the 3D pancreatic surface. Manual outlining will be performed by a clinical radiologist-trained fellow. A non-linear mixed effects (NLME) modeling approach will be employed to describe quantitatively the longitudinal changes of the pancreatic volume, enabling assessment of inter-individual and intra-individual variabilities.
Background Rationale
Currently, the pancreas imaging analysis is based on a mixture of qualitative analysis performed by well-trained pathologists for histopathology and manual tracing of the structures for further quantitative analysis by clinical radiologists. These methods performed manually require long processing times along with unavoidable human error factors and limit through-put. We will develop DL-based segmentation and detection algorithms tracing the boundaries of the pancreatic islets, identifying β-cells and CD3+T cells on immunohistochemistry WSI, and pancreatic volumes by MRI. These automated quantitative analysis tools will accelerate the analysis of pancreas images and provide insights into better understanding the heterogeneity of the pancreatic islets, which will help comparison across the different T1D state groups. These tools will be made available to the broader diabetes community as open source programs.
Description of Project
There is currently no approved therapy to prevent or delay the onset of type 1 diabetes (T1D) and there is a lack of biomarkers to identify individuals and quantify their risk of conversion to a diagnosis of T1D except for those with first-degree relatives with T1D. There have been significant failures in late-stage development of therapies for patients with new-onset T1D. These failures have been attributed in part to two factors. First, there is a high-degree of heterogeneity in age of onset in this patient population and a current inability to quantitatively describe the contributions of specific sources of variability to such heterogeneity. Second, intervening in new-onset T1D may be too late to significantly delay or halt disease progression and preserve endogenous β-cell function. T1D remains a major cause of suffering, a burden to society, and its incidence appears to be increasing, especially in younger children. In fact, by 2050 the number of individuals diagnosed with T1D in the US alone is projected to more than triple.
T1D is an autoimmune disease that results from the destruction of pancreatic islet β-cells. Islets represent well-organized micro-neuroendocrine organs whose architecture provides for maximally effective endocrine hormone secretion in response to a meal. Islets in patients with T1D by histology and immunolocalization, made possible through the JDRF Network for Pancreatic Organ donors with Diabetes (nPOD), display a high degree of heterogeneity in sizes, numbers, architecture, endocrine and immune cell type compositions, and capillary density, compared to non-diabetic controls. Clinical studies show that both the duration (months to years) and progression rate (rapid, slow) can be highly variable during the asymptomatic phases depending in part on the subject’s age (Stages 1-2). Younger subjects generally have a more rapid progression (months to 1-2 years) while older individuals may progress over many years. Notably, islets in organ donors with multiple islet autoantibodies show heterogeneous islet pathology in both numbers of islets with inflammation (insulitis) alongside a combination of observations noting islets with normal proportions of β-cells (INS+) adjacent to those islets with complete loss of β-cells (INS-). Importantly, both islet types can possess insulitis and numbers of such islets have been inadequately studied in non-diabetic controls or single autoantibody-positive donors. In addition, a small pancreas size in T1D was affirmed by the JDRF nPOD study using organ weights and by a NIH DP3 TrialNet study using MRI assessed volume; both studies noting size/volume was decreased ~45% at the time of T1D onset with minimal changes in pancreas size after onset. The small pancreas size infers the potential for exocrine abnormalities and perhaps cross-talk between the exocrine and endocrine compartments particularly as the endocrine compartment comprises only ~2% of the entire pancreas volume.
We propose to develop automated deep learning-based pancreatic islet imaging analysis and informatics tools using nPOD whole slide data and MRI data from the NIH DP3 TrialNet and on-going R01 studies, which will inform clinical trial designs evaluating the efficacy of potential therapies for T1D prevention and treatment.
Anticipated Outcome
We will develop deep learning-based segmentation and detection algorithms tracing the boundaries of the pancreatic islets and identifying β-cells, α-cells, and CD3+T cells (endocrine and exocrine), automatically and in hundreds of images representing all regions of the pancreas. To analyze the results obtained from the deep learning algorithms, we will employ statistical methods to quantify the variations across different T1D disease state groups. Based on the trained and evaluated deep learning models, we will develop a graphical user interface (GUI) to improve usability of the tools to meet the needs of pathologists. Users will be able to upload image data obtained by their own, and the GUI will show graphically the automatically generated annotations of the different region of interests as well as summarize in plots and/or tables the features quantified by the deep learning model and statistical analysis. The users will be able to save the results obtained in the GUI with their samples in a friendly format (e.g. .jpeg and .xlsx). By performing a usability test, an informative step-by-step manual and/or a tutorial video will be also created. The tool will be made available to the broader diabetes community.
Relevance to T1D
This automated quantitative analysis tool will accelerate the analysis of pancreas images and provide insights into better understanding the heterogeneity among T1D patient population. We also believe these data will assist with the analysis of morphological details using MRI or CT radiology for exocrine atrophy (reduced exocrine cell numbers), and interlobular and intralobular fibrosis and inflammation. As a long-term potential, the proposed tool can be extended and applied to extensive longitudinal imaging data. The pancreatic architecture-related features quantified by the AI-based tool would be new promising endpoints to model in order to develop a quantitative disease-drug-trial model-based clinical trial simulation tool, by applying the modeling and simulation approach that we are employing for the ongoing JDRF-funded project (2-SRA-2020-903-A-N). Ultimately, our results will help in the rational design of T1D prevention and treatment strategies.