Advancements in diagnosing oral potentially malignant disorders: leveraging Vision transformers for multi-class detection

Abstract

Objectives: Diagnosing oral potentially malignant disorders (OPMD) is critical to prevent oral cancer. This study aims to automatically detect and classify the most common pre-malignant oral lesions, such as leukoplakia and oral lichen planus (OLP), and distinguish them from oral squamous cell carcinomas (OSCC) and healthy oral mucosa on clinical photographs using vision transformers.

Methods: 4,161 photographs of healthy mucosa, leukoplakia, OLP, and OSCC were included. Findings were annotated pixel-wise and reviewed by three clinicians. The photographs were divided into 3,337 for training and validation and 824 for testing. The training and validation images were further divided into five folds with stratification. A Mask R-CNN with a Swin Transformer was trained five times with cross-validation, and the held-out test split was used to evaluate the model performance. The precision, F1-score, sensitivity, specificity, and accuracy were calculated. The area under the receiver operating characteristics curve (AUC) and the confusion matrix of the most effective model were presented.

Results: The detection of OSCC with the employed model yielded an F1 of 0.852 and AUC of 0.974. The detection of OLP had an F1 of 0.825 and AUC of 0.948. For leukoplakia the F1 was 0.796 and the AUC was 0.938.

Conclusions: OSCC were effectively detected with the employed model, whereas the detection of OLP and leukoplakia was moderately effective.

Clinical relevance: Oral cancer is often detected in advanced stages. The demonstrated technology may support the detection and observation of OPMD to lower the disease burden and identify malignant oral cavity lesions earlier.

Keywords: Artificial Intelligence; Deep learning; Leukoplakia; Malignant transformation; Oral lichen planus; Oral squamous cell carcinoma.

Fig. 1

Examples of correct OSCC, leukoplakia, and OLP predictions. The left column is the input image, the middle column is the reference annotation, and the right column is the prediction

Fig. 2

Confusion matrix of the results of multi-class image-level disorder detection. The results of the most effective cross-validation model on the test images are shown. Predicted disorders with a maximum score ≥ 0.75 are matched to reference disorder. The colors are normalized by the number of predicted labels

Fig. 3

Examples of incorrect predictions. The left column is the input image, the middle is the reference annotation, and the right is the prediction. The scores in the right column are the confidences of the model. The first row illustrates the false-positive prediction. The second row shows the false-negative prediction. The last row represents model’s misclassification

Fig. 4

Receiver operating characteristic (ROC) curves of image-level classification. On the left side, the ROC curve illustrates the binary classification (pathology versus no disorder). On the right side, the ROC curve shows the binary classifications for the respective pathologies (OSCC versus no disorder; leukoplakia versus no disorder; OLP versus no disorder). Each center line and peripheral line represent the mean and the mean plus and minus the standard deviation across cross-validation models