Deep Learning-based Analysis of Oral Cancer

Abstract

Can smartphone imaging hardware act as a substitute for WSI scanners in automated computational pathology? How much (if any) performance degradation is observed in deep learning algorithms when they are trained on smartphone acquired images instead of WSI scanned images. Slides from 12 cases of Oral Squamous Cell Carcinoma (OSCC) were scanned using both a Leica/Aperio CS2 WSI scanner and a smartphone placed at the eyepiece of a clinical microscope. Three features: Mitotic figures, Keratin Pearls and Epithelial regions were then labelled by expert pathologists in both WSI and smartphone acquired images. Deep learning algorithms were then trained and evaluated on both types of images for performance comparison. The largest degradation in performance was observed in mitosis detection (smartphone F1 score = 0.6437 Vs. WSI F1 score = 0.7059). In Keratin pearl detection, algorithms trained on smartphone images outperformed algorithms trained on WSI images (smartphone F1 score = 0.7312 Vs. WSI F1 score = 0.6984). Performance degradation was minor for Epithelium segmentation (smartphone F1 score = 0.7631 Vs. WSI F1 score = 0.7846).