Deep learning-based pathomics reveals renal tubular atrophy with inflammatory cell infiltration as a prognostic marker in IgA nephropathy

IgA nephropathy (IgAN) is the most common primary glomerulonephritis worldwide. However, there are still many limitations in relying solely on renal pathology to assess the risk IgAN progression. As an emerging technology, pathomics can assist renal pathologists in discovering subvisual features, thereby becoming an important tool for assessing the risk of IgAN progression. To utilize pathomics to identify valuable subvisual features in renal pathology for the prognostic assessment of IgAN, and to further validate the clinical value of these features. In this study, we selected 285 patients diagnosed with primary IgAN who underwent renal biopsy at Zhejiang Provincial Hospital of Chinese Medicine, Zhejiang Provincial Zhongshan Hospital, Zhejiang Provincial Dongfang Hospital, and Jiangshan Hospital of Chinese Medicine between January 1, 2017, and January 1, 2020. Among them, 200 patients were randomly allocated to train the deep learning model, while the remaining 85 were reserved for validation. Whole-slide images (WSI) of HE, PAS, Masson, and PASM staining from these patients were divided into 512 × 512 pixel patches at magnifications of 100×, 200×, and 400×. After removing blank patches, the Vahadane method was applied to standardize the color of all patches. The Residual Network 18 was utilized for deep learning to predict the risk of each patch experiencing the study outcome, and then the prediction probabilities of each patch were aggregated onto the corresponding WSI using the multi-instance learning. Principal component analysis was then applied to reduce the dimension of the data, and the optimal magnification rate was determined based on the C-index of the Cox model. Visualization techniques were further employed to project the prediction of the patch on the optimal condition onto the WSI. Finally, renal pathologists reviewed and summarized all the patches to identify the pathological features with prognostic value, and another pathologist independently validated these features discovered by the pathomics in the validation set. The prognostic model constructed using deep learning (0.837 ± 0.051) significantly outperformed the Lee grading (0.676 ± 0.059) and Oxford classification (0.766 ± 0.055) in predicting the occurrence of the adverse study outcome in patients with IgAN. Moreover, the key pathological feature identified by pathomics mainly focused on ‘renal tubular atrophy with inflammatory cell infiltration,’ which was validated as a valuable prognostic predictor in patients with IgAN through independent assessment by renal pathologists in the validation set. This study utilized pathomics to identify renal tubular atrophy with inflammatory cell infiltration as a feature associated with the progression risk of IgAN. This feature may complement the Oxford Classification (MEST-CL), thereby potentially enabling more accurate prognostic assessment in patients with IgAN.