Ensemble Deep Learning Framework for Oral Cancer Detection

Abstract

The early diagnosis of oral cancer is critical when it comes to enhancing the treatment process as well as increasing chances of survival. Unfortunately, there are challenges in designing automated classifiers since clinical image datasets with sufficient annotations can be scarce. In this work, an ensemble deep learning classifier is proposed that performs classification based on lip and tongue images from the perspective of early oral cancer screening. For dealing with possible data deficiency, methods such as data augmentation and synthesis were used. Specifically, three pre-trained models, namely ResNet-50, VGG19, and EfficientNet, are employed to extract complementary feature representations for accurate classification. The proposed framework achieved classification accuracies of 97.20% and 92.00% on the augmented and synthetic datasets, respectively.To improve model interpretability, Gradient-weighted Class Activation Mapping++ (Grad-CAM++) was used to generate attention maps. The obtained results prove the effectiveness of the designed method for solving the considered problem despite data deficiency challenges. Thus, this approach can be used as a computer-aided diagnosis system.