Explainability of a Deep Learning Based Classification Model for ANCA-Associated Glomerulonephritis

Journal of the American Society of Nephrology, vol. 33, pp. 32–33

Abstract

Background: The histopathological classification for ANCA-associated glomerulonephritis (ANCA-GN) is a well-established tool to reflect the variety of patterns and severity of lesions that can occur in renal biopsies of patients with ANCA-associated vasculitis. As with many fields, medicine saw a rapid emergence of Artificial Intelligence (AI) and Deep Learning (DL) approaches. In the field of digital pathology, AI can now serve as decision-support for pathologists, with the potential for gains in productivity and time-saving. It was demonstrated previously that AI can aid in identifying histopathological classes of renal diseases, e.g. of diabetic nephropathy. Although these models reach high prediction accuracies, their black box structure makes them very non-transparent. The disadvantage is that the networks' decisions are not easily interpretable by humans and it is not clear what information in the input data underlies their decisions. This necessitates the use of Explainable AI (XAI), so that decisions made by AI models become accessible for validation by a human expert. Methods: Renal biopsy slides of 80 patients with ANCA-GN from 3 European centers, who underwent a diagnostic renal biopsy between 1991 and 2011, were included. On the scanned slides glomeruli were labelled as ‘normal’, ‘sclerotic’, ‘crescentic’ or ‘abnormal - other’. We developed a DL-based computational pipeline, which detects and classifies the glomeruli. We investigated the explainability of our model, using XAI techniques to shed light on the decision-making criteria of our trained DL classifier, using saliency maps. These maps were analyzed by pathologists to compare the decision-making criteria of humans and the DL model. Results: Our DL model shows a prediction accuracy of 93% for classifying glomeruli. The saliency maps from our trained DL models help us to better understand the decision-making criteria of the DL black box. Conclusions: AI and DL play an increasingly important role in (nephro) pathology. To ultimately enable safe implementation of these models in clinical practice, validation of their decisions is needed. To achieve this, we used XAI techniques, which showed great potential for illuminating the decision-making criteria of the DL black box.

Authors 7

1. Maria A.C. Wester Trejo Aachen Department of Pathology, Leiden

   Leiden University Medical Center

   Affiliation as printed

   Leiden University Medical Center, Department of Pathology, Leiden, Netherlands

2. Maryam Sadeghi

   Innsbruck Medical University · Universität Innsbruck

   Affiliation as printed

   Medical University of Innsbruck, Department of Medical Statistics, Informatics and Health Economics, Innsbruck, Austria

3. Shivam Singh

   Innsbruck Medical University · Universität Innsbruck

   Affiliation as printed

   Medical University of Innsbruck, Department of Medical Statistics, Informatics and Health Economics, Innsbruck, Austria

4. Andreas Kronbichler

   University of Cambridge

   Affiliation as printed

   University of Cambridge, Department of Medicine, Cambridge, United Kingdom

5. Vladimı́r Tesař

   Charles University · General University Hospital in Prague

   Affiliation as printed

   Charles University and General University Hospital, Department of Nephrology, Prague, Czechia

6. Xavier Puéchal

   Hôpital Cochin

   Affiliation as printed

   Hôpital Cochin, Department of Internal Medicine, Paris, France

7. Ingeborg M. Bajema Aachen Department of Pathology, Leiden

   Leiden University Medical Center

   Affiliation as printed

   Leiden University Medical Center, Department of Pathology, Leiden, Netherlands