Artificial Intelligence-based Techniques to Characterize KIdney Microstructure on Histological ImagEs

Overview

The primary aim of this observational exploratory study will be to use fully anonymized histological images of kidney human tissue from patients with any kidney disease and normal kidney tissue to develop novel deep learning-based image processing techniques allowing to characterize kidney microstructure across different pathologies and/or disease stages. Secondly, the study will aim at validating the novel techniques against gold standard (manual) methods, when available, and at developing novel histological imaging biomarkers that could support differential diagnosis, staging of the disease, monitoring of disease progression and response to therapy, and prediction of the disease progression. Other exploratory aims will include: * The use of radiomics techniques to identify disease-specific kidney morphology patterns. * The implementation of uncertainty quantification techniques, able to increase AI explainability.

Morphology-based histopathological analysis of kidney tissue plays a key role in the diagnosis and therapeutic decisions of many kidney diseases. To date, histopathological analysis is mainly performed qualitatively, by visual inspection, requiring highly trained expert pathologists. Histopathologic findings are often scored by pathologists using semiquantitative diagnostic classification scales, such as the Oxford classification of IgA nephropathy, or disease severity scales. Despite such scoring systems, histopathological analysis remains semi-quantitative, time consuming, and highly operator-dependent. Manual techniques have been proposed to quantitatively assess kidney microstructure on histological images, showing potential to monitor disease progression and response to therapy in chronic kidney disease (CKD). As an example, peritubular interstitial volume, responsible for crucial endocrine functions and undergoing significant, albeit reversible, expansion in CKD, has been recently quantified on kidney biopsy specimens by point counting on each frame. Despite allowing accurate quantification, these manual techniques are labour-intensive and operator dependent. Fast and objective quantitative assessment of kidney microstructure would be highly desirable. The digitalisation of histological images, same as for diagnostic images, has made it possible to benefit from advanced image analysis techniques allowing identification and segmentation of relevant histopathological structures, and quantitative assessment of tissue microstructure. In the recent years, Artificial Intelligence (AI) and, in particular, Deep Learning (DL) techniques have shown promise for (semi)automated segmentation of relevant morphological structures on histological images, limiting the need for expert operators, ensuring reproducibility and massively reducing the time demand. Convolutional neural networks (CNNs) have recently demonstrated outstanding performance in image segmentation tasks, also in the medical field. In particular, the so-called U-Nets, consisting of a contracting and an expanding path, have become increasingly popular since first used. Few studies, so far, have used CNNs to investigate kidney microstructure on histological images. Hermsen et al. used CNNs for multi-class segmentation of histological images from kidney biopsies \[8\]. A similar study aimed to develop a CNN for segmentation of mouse renal tissue structures, such as glomeruli, tubules, arteries, and veins, based on densely annotated images from different renal diseases and various animal species. Despite these promising preliminary efforts, the high heterogeneity of morphological patterns poses challenges to the generalizability of the segmentation techniques. Automated DL-based methods able to accurately segment and quantify relevant morphological structures on histological kidney images from patients with different kidney pathologies and/or different disease stage would be highly desirable.