Explainable deep transferlearning approach for classifying oral tongue lesions

Abstract

Oral cancer has recently become a prevalent disease worldwide, with the highest reported cases in South Asian region. Due to the heterogeneous features of oral tongue cancers, receiving medical attention only after spreading to neighboring areas is inevitable. Early identification is vital to increase the survival rate of the patients and for less expensive post-care after the treatments. However, conventional screening methods have known limitations that prevent them from diagnosing pre-cancerous lesions at their early stages. Hence, researchers study alternative screening methods and point-of-care testing tools that utilize cutting-edge technologies to overcome those limitations in current diagnostic techniques.

Although deep learning techniques have shown propitious results in diagnosing numerous types of cancer using various medical images, research focused on oral tongue lesions are limited. To perform exceptionally, these models need large amounts of data. However, due to the privacy concerns of patients, gathering and storing such a vast dataset can be onerous. Thus, the primary aim of the study is to implement a method that can learn from a limited number of publicly available images, automate the identification of a potential presence of oral tongue cancer, and locate the features used in classification to make the model explainable.

Among the models trained, InceptionV3 and MobileNetV2 demonstrated the highest accuracy with 78% and 75%, respectively. However, the Grad-CAM heatmaps of the MobileNetV2 model indicated the most correct regions of interest in the images with 70% of average intersection over union value, compared to the other models.