Aprendizado profundo aplicado à quantificação automática de variáveis morfométricas depodócitos em imagens de imuno-histoquímica

Abstract

Among the various cell types that make up renal tissue are podocytes, highly specialized glomerulus cells that surround the capillaries and neighboring Bowman capsule cells. Genetic or acquired involvement of podocytes can lead to the development of various proteinuric diseases. The hypothesis of podocyte depletion has gained considerable attention in the last decade, mainly because it represents a unifying concept in renal pathology which associates this condition with the development of several glomerular diseases, especially those characterized by glomerulosclerosis, which evolve to chronic kidney disease. Therefore, several methodologies for quantitative evaluation of podocytes by simpler and widely available means have been developed. The objectives of the present study were: 1) to verify the use of immunohistochemistry (IH) as an alternative marker to immunofluorescence (IF) to quantify morphometric parameters, especially podocyte density, in renal biopsies by the correction factor method (VENKATAREDDY et al., 2014) and 2) present an automated evaluation for quantification of these attributes from a Convolutional Neural Network. Sixteen pre-transplant renal biopsy cases from deceased donors from 2016 to 2018 were studied. Samples were labeled with IH by anti-WT1 nuclear antibody (wilm’s tumor type 1). This marking was able to present podocyte density results similar to those reported in the literature through IF, with a coefficient of variation between studies of 14%. IH with anti-WT1 antibody was found to have a broad spectrum labeling of the pixels belonging to the nucleus images, characterized by large pigmentation variability between individuals, between glomeruli of the same individual and within the same glomerulus. This factor added difficulty and more time to the process of recognition by the professional. In the development of the Convolutional Neural Network (U-net: Convolutional Networks for Biomedical Image Segmentation), 122 glomerular images with 2,977 podocyte nucleus images were used. Where feature vectors were extracted from 80% of the randomly selected images for Machine Learning model training and the remaining 20% were used in the Neural Network to obtain the desired metrics. The construction of the Neural Network was based on the best performance criteria observed regarding the sensitivity and precision in the recognition of pixels belonging to podocyte nucleus images. All 6 morphological variables tested pairwise showed no statistical differences between the manual podometry process and the use of the Neural Network. The similarity between binary markers and neural network masks was 89%, as measured by the Jaccard Coefficient. It was concluded that the Neural Network presented a relevant performance for podocyte nucleus identification and counting, as well as for the measurement of nuclear diameters. In addition, there was a significant advantage in optimizing the manual process by reducing the time and effort spent on this activity. Considering that the Machine Learning process can be progressively improved by inserting more information into the Neural Network, a great potential of this model is observed as it is fed with more images and / or use of more specific biomarkers.