D’OraCa: Deep Learning-Based Classification of Oral Lesions with Mouth Landmark Guidance for Early Detection of Oral Cancer

Abstract

Oral cancer is a major health issue among low- and middle-income countries due to the late diagnosis. Automated algorithms and tools have the potential to identify oral lesions for early detection of oral cancer. In this paper, we aim to develop a novel deep learning framework named D’OraCa to classify oral lesions using photographic images. We are the first to develop a mouth landmark detection model for the oral images and incorporate it into the oral lesion classification model as a guidance to improve the classification accuracy. We evaluate the performance of five different deep convolutional neural networks and MobileNetV2 was chosen as the feature extractor for our proposed mouth landmark detection model. Quantitative and qualitative results demonstrate the effectiveness of the mouth landmark detection model in guiding the classification model to classify the oral lesions into four different referral decision classes. We train our proposed mouth landmark model on a combination of five datasets, containing 221,565 images. Then, we train and evaluate our proposed classification model with mouth landmark guidance using 2,455 oral images. The results are consistent with clinicians and the F1 score of the classification model is improved to 61.68%.