Automated Detection and Classification of Oral Pre-malignant and Malignant Lesions Using AI Techniques

Abstract

Oral cancer is one of the most prevalent cancers in the world and has a high mortality rate, primarily as a result of late diagnosis. Early identification of premalignant lesions such as leukoplakia and erythroplakia can greatly increase the survival rate; however, traditional diagnostic techniques are based on visual perception and biopsy examination, which are subjective, invasive, and time consuming. This work presents a fully automated AI system for the detection and classification of oral premalignant and malignant lesions. A set of clinical oral lesion images were gathered from publicly available medical repositories covering precancerous lesions and malignant types e.g., Oral Squamous Cell Carcinoma (OSCC), verrucous carcinoma, sarcoma, lymphoma, malignant melanoma, metastatic lesions. In order to handle data deficiency and class imbalance, we used several data augmentation methodologies such as rotation, horizontal flipping, brightness and contrast adjustment, scaling and Gaussian blur. The images were pre-processed by resizing, normalization, noise filtering and contrast stretching. The convolutional neural network model EfficientNet-B0 was used for feature extraction and the similarity-based k-Nearest Neighbors (k-NN) is employed for classification. An ensemble strategy was further employed by integrating complementary feature representations to improve classification robustness and accuracy. The highest score for predicted categories of lesions are displayed and then saved to a local database.