Automatic Evaluation of Histological Prognostic Factors Using Two Consecutive Convolutional Neural Networks on Kidney Samples

Abstract

Background and objectives: The prognosis of patients undergoing kidney tumor resection or kidney donation is linked to many histologic criteria. These criteria notably include glomerular density, glomerular volume, vascular luminal stenosis, and severity of interstitial fibrosis/tubular atrophy. Automated measurements through a deep-learning approach could save time and provide more precise data. This work aimed to develop a free tool to automatically obtain kidney histologic prognostic features.

Design, setting, participants, & measurements: In total, 241 samples of healthy kidney tissue were split into three independent cohorts. The "Training" cohort (n=65) was used to train two convolutional neural networks: one to detect the cortex and a second to segment the kidney structures. The "Test" cohort (n=50) assessed their performance by comparing manually outlined regions of interest to predicted ones. The "Application" cohort (n=126) compared prognostic histologic data obtained manually or through the algorithm on the basis of the combination of the two convolutional neural networks.

Results: In the Test cohort, the networks isolated the cortex and segmented the elements of interest with good performances (>90% of the cortex, healthy tubules, glomeruli, and even globally sclerotic glomeruli were detected). In the Application cohort, the expected and predicted prognostic data were significantly correlated. The correlation coefficients r were 0.85 for glomerular volume, 0.51 for glomerular density, 0.75 for interstitial fibrosis, 0.71 for tubular atrophy, and 0.73 for vascular intimal thickness, respectively. The algorithm had a good ability to predict significant (>25%) tubular atrophy and interstitial fibrosis level (receiver operator characteristic curve with an area under the curve, 0.92 and 0.91, respectively) or a significant vascular luminal stenosis (>50%) (area under the curve, 0.85).

Conclusion: This freely available tool enables the automated segmentation of kidney tissue to obtain prognostic histologic data in a fast, objective, reliable, and reproducible way.

Keywords: computer; deep learning; neural networks; prognosis; renal pathology.

Copyright © 2022 by the American Society of Nephrology.

Figures

Graphical abstract

Figure 1.

Flow chart of the Training/validation, Test, and Application cohorts used to develop the convolutional neural networks. ROI, region of interest; MCD, minimal change disease.

Figure 2.

Kidney samples stained with Masson’s trichrome before and after convolutional neural network predictions. A region of interest (A) with the neural network prediction (B) (×200 magnification). False positives were observed with a peritubular capillary wrongly recognized as a vein (*) and interstitial tissue as an artery (¥). Internal and external elastic laminas were wrongly overexpanded in the upper vessel. A kidney biopsy of a patient with a minimal change disease (C) with the prediction of the cortical area (D) and smaller kidney elements (E) (×10 magnification). (B), (E) normal tubules (red), atrophic tubules (orange), Bowman’s capsule (yellow), nonsclerotic glomeruli (light green), and globally sclerotic glomeruli (light blue), internal elastic lamina (pink), external elastic lamina (purple), vein (deep blue). (D) Cortical area (red), capsule (deep blue), medullary area (green).

Figure 3.

Confusion matrix per pixel to assess the performance of the Mask R-CNN neural network for multiclass segmentation in the Test cohort. For example, for pixels belonging to the category of nonsclerotic glomeruli (pixels having been manually assigned to this category), the neural network correctly predicted the right category for 93% of those pixels and predicted a wrong one (interstitium) for 7% of those pixels. The Matthews correlation coefficient was 0.85 for this confusion matrix.

Figure 4.

Evaluation of prognostic factors in the Application cohort. The factors assessed were mean glomerular volume (A), glomerular density (B), interstitial fibrosis (C–E), tubular atrophy (F–H), and vascular luminal stenosis through intimal thickening (I–K). The significant correlations between the observed factors and those predicted by the convolutional neural networks were assessed by Pearson or Spearman correlation tests on the basis of whether the distribution was normal (A–C), (F), (I). Bland–Altman plot showing a systematic overestimation of predicted tubular atrophy or interstitial fibrosis (D), (G) and an underestimation of luminal stenosis (J). The mean bias is represented by the big dashed black lines with the 95% limits of agreement represented by the small dashed lines. Receiver operating characteristic curves assessing the capacity of the algorithm to predict interstitial fibrosis, and tubular atrophy over 25% (E), (H) and a luminal stenosis over 50% (K).