On the Influence of Beta Cell Granule Counting for Classification in Type 1 Diabetes

Patients suffering with type 1 diabetes show a major reduction of β -cells within their pancreas. By analyzing tissue samples containing granules – the insulin-producing units within the β -cells – we aim to gather more information on the respective healthy and diabetic phenotypes, which could lead to further understanding the pathogenesis of the disease. To this end, we use a deep learning approach to investigate whether assumptions on the pathological status can be made based on electron micrograph images of β -cells. To support the decision- making process we explore whether estimating the number of granules can be used to aid in discriminating healthy from diabetic samples. Furthermore, we demonstrate that multi-task and transfer learning strategies can lead to more accurate predictions. Finally, this work intends to contribute to a more in-depth understanding of the structural mechanisms in type 1 diabetes, which is essential to design better approaches to a tailored treatment.