Deep Learning of Pancreas MRI to Predict Progression of T1D.

点击翻译按钮获取中文摘要

Project Summary The TrialNet Pathway to Prevention study has provided crucial early screening for relatives of individuals with type 1 diabetes (T1D). The presence of autoantibodies conveys high risk for progression from Stage 1 T1D, defined by the presence of multiple diabetogenic autoantibodies, to Stage 3 T1D, or symptomatic disease. However, the time to progression can be variable. A variety of genetic and metabolic indices have attempted to predict progression of T1D, with varying degree of success. Additional biomarkers are needed to improve prediction of progression, and these biomarkers must be correlated with immunological markers or metrics that assess beta cell function. The overall goal of the proposed study is to establish an imaging biomarker to predict progression. We propose to improve T1D prediction by 1) co-registering longitudinal MRI taken during progression of T1D to identify spatial evolution characteristic of disease evolution, 2) harnessing deep learning techniques to identify image features characteristic of the pancreas in T1D, and 3) integrated imaging and functional metrics to build a predictive model of T1D progression. This work builds upon work we have performed indicating that pancreas size, shape, and structure are altered in new onset type 1 diabetes. These imaging metrics are also altered in individuals at risk for developing T1D. This study will identify imaging features characteristic of the pancreas that accompany progression to T1D. The techniques developed may prove useful for monitoring patients at risk for T1D and predicting progression to symptomatic disease, which is associated with lower incidence of diabetic ketoacidosis at diagnosis, better glycemic control, and corresponding improvements in long-term complications. The ability to predict progression would further facilitate the design of new therapeutic trials which are shorter and less expensive by stratifying patient populations and providing intermediate end points.