#6695 THE ROLE OF A DEEP LEARING MODEL FOR EVALUATING HISTOPATHOLOGY OF ACUTE RENAL TUBULAR INJURY

Abstract

Abstract Background and Aims Acute tubular injury is the most common intrinsic cause in acute kidney injury. Histopathological diagnosis may help distinguish between different types of acute kidney injury and aid in treatment. Currently, there have not been any studies that uses deep learning model for assisting histopathological diagnosis in acute kidney injury. The aim of our study is histopathological classification and segmentation identified four structures of acute renal tubular injury based on deep learning models. Method We used injury classification model to evaluate the degree of renal injury (grade 0, grade 1 and grade 2) and segmentation model (glomerulus, healthy tubules, tubules with cast, and necrotic tubules) specifically to classify the tubular specific injury after cisplatin treatment. We trained a classification model using ResNet-18 to classify the grade of acute tubular injury and a segmentation model using DeepLabV3 with ResNet-50 backbone for four structures of the mouse PAS-stained histopathology of acute renal tubular injury. Results The classification algorithm achieved area under the receiver operating characteristics curve of 0.98, 0.77, and 0.80 for grade 0, grade 1, and grade 2, respectively. In segmentation model for four structures, the highest Intersection over Union and the Dice coefficient were obtained for the segmentation of “glomerulus” class, followed by the “healthy tubules”, “tubules with cast”, and “necrotic tubules” classes. The overall performance of the segmentation algorithm for all classes in the test set is Intersection over Union 0.7256, Dice coefficient 0.8185. Conclusion The predictive model by deep learning could identify the histopathological structures and classify the grade of renal tubular injury, and may contribute to a more effective evaluation of renal pathology.