Early Tongue Cancer Detection in Photographs Using a Pretrained Convolutional Neural Network

Although the tongue is an easily visible organ, tongue cancer often goes undetected until an advanced stage because it is difficult to distinguish between malignant lesions and non-malignant lesions. This study assessed the detection potential of tongue cancer, precancerous lesions, and inflammatory lesions using a pretrained convolutional neural network (CNN) and evaluated its effectiveness in a situation with multiple classes and imbalanced datasets. In total, 1,810 tongue images, each carefully labeled by medical specialists, were used for model training. Transfer learning, data augmentation, and fine-tuning were used to overcome the problems associated with limited datasets. Also, the weight balancing method was introduced to mitigate class imbalance. Three popular pretrained CNNs, namely VGG16, Inception-ResNet-V2, and EfficientNet, were evaluated as a backbone network. The final model achieved an accuracy of 0.9167, a precision of 0.9212, a recall of 0.9167, and an F 2 score of 0.9176. Our results show that a pretrained CNN with a moderate complexity and a deep architecture based on data bypassing can detect and differentiate tongue lesions by applying currently available deep learning techniques, even with multiple classes and a limited and disproportionate number of images. Thus, these strategies can facilitate timely tongue lesion diagnosis and prompt treatment.