FC046: Automated Mest-C Classification in IGA Nephropathy using Deep-Learning based Segmentation

Abstract

Abstract BACKGROUND AND AIMS IgA nephropathy prognosis depends on histological factors. The MEST-C score is used to grade those factors and assess the renal prognosis of this disease. Nevertheless, this manual evaluation is time-consuming, tedious and has a poor reproducibility. An automated analysis would be faster and more objective. This work aimed to use deep-learning techniques on whole kidney biopsies from patients suffering from IgA nephropathy to obtain an automated MEST-C score. METHOD We used a previously developed convolutional neural network (CNN) to isolate the cortical area. Then, two additional CNNs were independently trained. The first one was used to evaluate the Interstitial Fibrosis/Tubular Atrophy (IF/TA) and segment the glomeruli. The second one was used to segment the relevant glomerular lesions (mesangial hypercellularity, endocapillary hypercellularity, segmental glomerulosclerosis and crescents). A total of 95 kidney biopsies from patients suffering from IgA nephropathy were randomly sorted into either the training set or the validation set. From those samples, regions of interest (ROI) were selected and annotated to develop the two CNNs (358 ROI and 458 ROI, respectively). The main annotation categories for the first CNN were `non-sclerotic glomeruli’, `sclerotic glomeruli’, `normal tubules’, `atrophic tubules’, `veins’ and `arteries’. For the second CNN, the main categories were `mesangial hypercellularity’, `endocapillary hypercellularity’, `segmental glomerulosclerosis’ and `crescents. Then, the three CNNs were sequentially applied to the biopsies of 109 patients suffering from IgA nephropathy. From the data provided by the CNNs, an automated MEST-C was calculated. We assessed its performances to predict the criteria of the gold standard MEST-C (scored by nephropathologists) using Receiver Operator Characteristic (ROC). RESULTS The CNNs detected the renal structures of interest with good precision and recall up to 0.98 and 0.96 respectively for the non-sclerotic glomeruli. Normal and atrophic were also reliably recognized with 0.88 and 0.71 f-scores. Glomerular lesions were harder to recognize with a 0.67 f-score for mesangial hypercellularity, 0.76 f-score for endocapillary hypercellularity, 0.47 f-score for segmental glomerulosclerosis and 0.50 f-score for the crescents. Manual and automated evaluation of IF/TA were well correlated with a 0.76 Spearman coefficient (P < 0.001). In the application cohort, the criteria of the automated adapted MEST-C were compared with the criteria of the visual MEST-C using Receiver Operator Characteristic (ROC), we assessed the performances of the criteria of the automated MEST-C to predict the criteria of the manual MEST-C. Areas under the curve (AUC) were 0.82 for mesangial hypercellularity, 0.79 for endothelial hypercellularity and 0.82 for segmental glomerulosclerosis. Regarding IF/TA (the T criteria), the AUC were 0.86 for T1 and 0.84 for T2. Regarding the evaluation of crescents, the AUC were 0.76 for the C1 criteria and 0.88 for the C2 criteria. Those ROC were used to determine adapted thresholds for each criterion. CONCLUSION This deep-learning based tool can be used to segment relevant renal cortical structures and provide a reliable assessment of IF/TA. An automated MEST-C score can be provided by our tool. Nevertheless, improvements are needed to consider using this tool in clinical practice and obtain an objective evaluation on large and numerous samples. FIGURE 1:Kidney samples stained with Masson's trichrome before and after the two first consecutive convolutional neural networks predictions (B, C). B: Cortical area (red), capsule (deep blue), C: normal and atrophic tubules (red and orange), non-sclerotic glomeruli (yellow), sclerotic glomeruli (light blue), arteries (purple), vein (deep blue).FIGURE 2:Kidney samples stained with Masson's trichrome before (A) and after the last convolutional neural networks predictions (B). B: hilum (yellow), mesangial hypercellularity (red), endothelial hypercellularity (purple), segmental glomerulosclerosis (green).