Detection and classification of glomerular lesions in kidney graft biopsies using 2-stage deep learning approach.

Acute allograft rejection in patients undergoing renal transplantation is diagnosed through histopathological analysis of renal graft biopsies, which can be used to quantify elementary lesions. However, quantification of elementary lesions requires considerable expertise, time, and effort. Using a 2-stage classification strategy, we sought to examine the effectiveness of deep learning in detecting and classifying glomeruli into 4 groups, namely normal, abnormal, sclerotic, and glomerulitis, as a potential biopsy triage system focused on transplant rejection. We used the U-Net model to build a glomeruli detection model using 137 kidney biopsy slides obtained from 80 kidney transplant patients. The median age of the patients was 52 (19-74) years, with 65% being men and 35% women. MobileNetV2 and VGG16 models were compared using a 2-stage classification strategy. In the first classification step, the models classified glomeruli into sclerotic and nonsclerotic glomeruli. In the second classification step, the nonsclerotic glomeruli from the first step were classified as normal, abnormal, or glomerulitis. The U-Net model achieved satisfactory detection (Dice coefficient = 0.90), and the MobileNetV2 model was the best for the 2 classification steps, with F1 scores of 0.85, 0.91, 0.98, and 0.92 for normal, abnormal, sclerotic, and glomerulitis, respectively. The 2-stage classification strategy identifies sclerotic glomeruli and abnormal glomeruli relative to permeable glomeruli and quantifies glomerulitis with significant accuracy while avoiding bias from abnormal glomeruli that do not have glomerulitis, providing valuable diagnostic information.