Deep learning algorithm for accurate diagnosis of oral cancer and precancers.

Abstract

e18135 Background: Oral cancer (OC) caused approximately 178,000 deaths in 2020 globally, often due to late stage diagnosis and poor survival rates. A proportion of OCs arise from precancers like leukoplakias. Early detection and accurate diagnosis are crucial, especially in low and middle income countries with limited access to specialized care. A chair-side non-invasive AI model capable of accurately detecting and classifying suspicious oral lesions could improve diagnostic accuracy and early detection. This study aimed to investigate the application of Deep Learning Algorithms for the automated classification of oral lesions using high-resolution clinical images. Methods: We collected intraoral images from patients seen at two large cancer centers (Sri Lanka and USA). Images were resized to 320 and clustered in four super classes: OC, Precancer, Benign and Normal. During training, random augmentations were applied to the images, including rotation, flipping, and exclusion of pixels. The YOLO (You Only Look Once) v11 architecture was selected, as it provides the highest performance in several state-of-the-art problems. The Classifier variant of the architecture was chosen in order to provide classification output. The algorithm was pre-trained on ImageNet. The training was done using an RTX 4090 video card using the AdamW optimiser with an initial learning rate of 0.000714 and momentum of 0.9, and a batch size of 64 for 35 epochs. Data analysis was performed using the confusion matrix, the average precision (positive predicted value), the average recall (sensitivity), and the overall accuracy. Results: We included 5,259 intraoral images (3.25 % OC, 40.96% pre-cancer, 39.46% benign and 16.33% normal mucosa). The YOLOv11-CLS algorithm achieved an average precision of 71%, an average recall of 55%, a specificity of 0.91%, and an overall accuracy of 62% on the validation set. On the test set, the model achieved an average precision of 60%, an average recall of 56%, a specificity of 87%, and overall accuracy of 65% on a test set. Performance on the OC class was low due to the very small number of OC cases included in the dataset. Conclusions: The deep learning-based classification showed promising signs of being able to classify suspicious oral cavity lesions and may therefore serve as a valuable tool to assist clinicians in early detection of OC/precancer, potentially improving patient outcomes through timely intervention. To improve the performance of the model, a larger dataset with a better class balance is required. Metrics on the validation and test sets. Class Precision (Positive Predicted Value) Sensitivity (Recall) Specificity Set Validation Test Validation Test Validation Test Benign 0.48 0.65 0.48 0.72 0.92 0.75 OC 0.96 0.53 0.57 0.31 0.99 0.99 Normal 0.86 0.62 0.44 0.57 0.96 0.93 Precancerous 0.57 0.68 0.73 0.65 0.80 0.79 Average 0.71 0.60 0.55 0.56 0.91 0.87