Diagnosis with Confidence: Deep Learning for Reliable Classification of Squamous Lesions of the Upper Aerodigestive Tract

Abstract

AbstractDiagnosis of head and neck squamous dysplasia and carcinomas is critical for patient care, cure and follow-up. It can be challenging, especially for intraepithelial lesions. Even though the last WHO classification simplified the grading of dysplasia with only two grades (except for oral or oropharyngeal lesions), the inter and intra-observer variability remains substantial, especially for non-specialized pathologists. In this study we investigated the potential of deep learning to assist the pathologist with automatic and reliable classification of head and neck squamous lesions following the 2022 WHO classification system for the hypopharynx, larynx, trachea and parapharyngeal space. We created, for the first time, a large scale database of histological samples intended for developing an automatic diagnostic tool. We developed and trained a weakly supervised model performing classification from whole slides images. A dual blind review was carried out to define a gold standard test set on which our model was able to classify lesions with high accuracy on every class (average AUC: 0.878 (95% CI: [0.834-0.918])). Finally, we defined a confidence score for the model predictions, which can be used to identify ambiguous or difficult cases. When the algorithm is applied as a screening tool, such cases can then be submitted to pathologists in priority. Our results demonstrate that the model, associated with confidence measurements, can help in the difficult task of classifying head and neck squamous lesions.