A Deep Learning Approach for Histopathological Diagnosis of Onychomycosis: Not Inferior to Analogue Diagnosis by Histopathologists

Abstract

Onychomycosis is common. Diagnosis can be confirmed by various methods; a commonly used method is the histological examination of nail clippings. A deep learning system was developed and its diagnostic accuracy compared with that of human experts. A dataset with annotations for fungal elements was used to train an artificial intelligence (AI) model. In a second dataset (n=199) the diagnostic accuracy of the AI was compared with that of dermatopathologists. The results show a non-inferiority of the deep learning system to that of analogue diagnosis (non-inferiority margin 5%) with respect to specificity and the area under the receiver operating characteristic curve (AUC). The AI achieved an AUC of 0.981. One limitation of this system is the need for a large number of training images. The AI had difficulty recognizing spores and confused serum or aggregated bacteria with fungal elements. Use of this deep learning system in dermatopathology routine might help to diagnose onychomycosis more efficiently.

Keywords: deep learning; dermatopathology; onychomycosis; artificial intelligence.

Fig. 1

Upper row: different patches with hyphae and their hyphae probabilities predicted by artificial intelligence (AI). Lower image: whole-slide image of the periodic acid–Schiff (PAS)-stained nail specimen. Red squares: location of numbered patches. Magnification: 400x

Fig. 2

Overview over artificial intelligence (AI) training and study. (a) A total of 528 periodic acid–Schiff-stained whole-slide images (WSI) were used for AI training. Markers for hyphae are shown as red dots (only for visualization of the annotations; red dots are not part of the actual training images). (b) Positive and negative patches were extracted for training from the WSI. (c) A convolutional neural network (CNN) was trained that predicts for a patch whether onychomycosis is present. (d) A total of 199 new WSI were used to study the performance of the AI in comparison with human experts. (e) To make an AI prediction for a WSI, the WSI was partitioned into patches. (f) The CNN is now fixed. (g) It predicts a probability for each patch independently. (h) All patch predictions were aggregated into a single probability estimate for the WSI.

Fig. 3

The receiver operating characteristic (ROC) curve shows that the onychomycosis probabilities predicted by artificial intelligence (AI) allow it to sort cases on an accuracy level comparable to that of human experts.

Fig. 4

Illustration of patches that were misclassified by artificial intelligence: (A–F) false-positives; (G–I) false-negatives. (A and B) Serum. (C) Overlay of nail fragments. (D and E) Bacteria in linear arrangement. (F) Remnants of neutrophils in zones of parakeratosis. (G and H) Hyphae were few and the area was not completely focused in the scan. (I) Hyphae were mainly crosscut. Magnification 400x.