Abstract

Onychomycosis is typically diagnosed via direct microscopy and culture. The most sensitive and reliable method for finding fungi is by staining with Periodic Acid Schiff. While organisms are often readily identifiable on PAS, in some biopsies fungal organisms may be scarce. These cases require extra time and effort by the pathologist and may lead to a false negative diagnosis. The aim of this study was to develop and validate a deep learning-based algorithm for the detection of fungal organisms on PAS-stained digital slides. Cases of nail biopsies (N=39) for suspected onychomycosis, diagnosed between 2019 and 2020 at our institution were selected and digitized using a Hamamatsu NanoZoomer S210 Scanner. A subset 20 WSI (16 positive, 4 negative) were selected for training the algorithm and the remaining 19 slides (17 positive, 2 negative) were reserved for model evaluation. QuPath v0.2.3 was used for digital slide annotation. A deep learning PAS detection algorithm was developed in Python 3.8 using TensorFlow 2.5.0. A MobileNetV2 neural network (CNN) architecture was implemented for the model. Performance was evaluated using a receiver operating characteristic curve (ROC curve) for each of three models developed for 10x, 20x, and 40x magnifications. The 40x model showed an AUC of 0.988 with a sensitivity and specificity of 0.979 and 0.866, respectively. The 20x model showed an AUC of 0.986, sensitivity of 0.98, and specificity of 0.843. The 10x model showed an AUC of 0.937, sensitivity of 0.982, and specificity of 0.476. We demonstrate that an AI model can be trained to automatically detect fungal organisms on PAS-stained digital slides. This can prove useful in practice as these algorithms can be used to screen cases before the pathologist sees them. For challenging cases in which rare fungi are present, these tools can potentially save the pathologist a significant amount of time and effort.

Financial Disclosure:
No current or relevant financial relationships exist.