Exploring the subtle and novel renal pathological changes in diabetic nephropathy using clustering analysis with deep learning.

To decrease the number of chronic kidney disease (CKD), early diagnosis of diabetic kidney disease is required. We performed invariant information clustering (IIC)-based clustering on glomerular images obtained from nephrectomized kidneys of patients with and without diabetes. We also used visualizing techniques (gradient-weighted class activation mapping (Grad-CAM) and generative adversarial networks (GAN)) to identify the novel and early pathological changes on light microscopy in diabetic nephropathy. Overall, 13,251 glomerular images (7,799 images from diabetes cases and 5,542 images from non-diabetes cases) obtained from 45 patients in Kanazawa Medical University were clustered into 10 clusters by IIC. Diabetic clusters that mainly contained glomerular images from diabetes cases (Clusters 0, 1, and 2) and non-diabetic clusters that mainly contained glomerular images from non-diabetes cases (Clusters 8 and 9) were distinguished in the t-distributed stochastic neighbor embedding (t-SNE) analysis. Grad-CAM demonstrated that the outer portions of glomerular capillaries in diabetic clusters had characteristic lesions. Cycle-GAN showed that compared to Bowman's space, smaller glomerular tufts was a characteristic lesion of diabetic clusters. These findings might be the subtle and novel pathological changes on light microscopy in diabetic nephropathy.