Prediction of the risk of cancer and the grade of dysplasia in leukoplakia lesions using deep learning.

Objectives: To estimate the probability of malignancy of an oral leukoplakia lesion using Deep Learning, in terms of evolution to cancer and high-risk dysplasia.

Materials And Methods: A total of 261 oral leukoplakia lesions with a mean of 5.5 years follow-up were analysed from standard digital photographs. A deep learning pipeline composed by a U-Net based segmentation of the lesion followed by a multi-task CNN classifier was used to predict the malignant transformation and the risk of dysplasia of the lesion. An explainability heatmap is constructed using LIME in order to interpret the decision of the model for each output.

Results: A Dice coefficient of 0.561 was achieved on the segmentation task. For the prediction of a malignant transformation, the model provided a sensitivity of 1 with a specificity of 0.692. For the prediction of high-risk dysplasia, the model achieved a specificity of 0.740 and a sensitivity of 0.928.

Conclusion: The proposed model using deep learning can be a helpful tool for predicting the possible malignant evolution of oral leukoplakias. The generated heatmap provides a high confidence on the output of the model and enables its interpretability.