Deep learning and pathomics augmented nephropathology

点击翻译按钮获取中文摘要

Nephropathology is essential for the diagnosis of renal diseases and a major read-out of experimental pre-clinical kidney studies. The advancement in digital pathology , i.e. digitalization of histological slides, and the application of artificial intelligence, especially deep learning, has opened new research perspectives that have the potential to transform diagnostic pathology into quantitative "computational" pathology. We aim to develop and apply deep learning in both clinical and experimental nephropathology. We have established a state-of-the-art high-throughput digital pathology infrastructure and platform which is further supported by our highly complementary, interdisciplinary expertise and long-term cooperation in this field. We aim to develop deep learning approaches to automatically segment various kidney compartments and perform classification among different domains including various animal species, human samples, stains, and diseases. We will address various challenges and limitations prevalent in digital pathology, including the unavailability of manually annotated data and staining variations, by developing semi- or un-supervised and stain-independent approaches. Additionally, we will perform exhaustive analyses to further our understanding of renal histopathology by the development of pathomics, i.e. large-scale extraction of quantitative image features that might detect previously unrecognized morphological attributes in each kidney compartment for different domains. As a proof-of-concept, we will use our trained networks to generate pathomics data on our experiments studying the role of desmosomes as biomarkers of kidney injury. Besides, we will provide our nephropathology and image analyses expertise for the whole consortium, strongly enhancing the comparability of the data generated while at the same time substantially increasing the datasets for deep learning developments. In conclusion, we aim to develop and apply deep learning and pathomics approaches to facilitate innovative quantitative pathology diagnostics towards a more precise and personalized nephropathology.