Nail clipping followed by periodic acideSchiff with diastase (PAS-D) staining has become the gold standard for diagnosis of onychomycosis, with relative ease of performance and high sensitivity. However studies have suggested this method may be less cost effective than inoffice alternatives due to the expense incurred by staining and pathologist interpretation. We trained a convolutional neural network (CNN) to classify digitized slides of nail clippings as positive or negative for onychomycosis. Specimens stained with both hematoxylin and eosin (H&E, n¼134) and PAS-D (n¼135) were digitized at 40X magnification and subsequently divided into training, validation, and test cohorts. Image preprocessing included tiling each composite image into 299 x 299 pixel patches and labeling non-empty patches as positive or negative based on prior diagnosis. Training was performed with the Inception-v3 CNN architecture with separate models developed for H&E-or PAS-D-stained slides. Initial classification of each stained specimen was obtained by averaging tile scores. For H&E-stained slides, using average tile score on test set specimens, the area under the curve (AUC) for the receiver operating characteristic curve was 0.86 (95% confidence interval [CI]: 0.74-0.96). For the model trained on PAS-D-stained slides, the analogous AUC was 0.88 (95% CI: 0.77-0.97). These results support the potential role of CNNs to automate diagnosis of onychomycosis, even on H&E-stained slides. Further studies to classify all imaged tissue per specimen rather than subcomponent tiles and to compare CNN performance to dermatopathologists are underway.

No abstract

No abstract

No abstract