Deep Learning Algorithm Spots Pre-Cancerous Tongue Lesions with High Accuracy

Oral cancer is a highly perilous condition, primarily because it is frequently identified only at advanced stages. Current screening methods often struggle with accuracy, missing the vital opportunity for early detection that could significantly reduce mortality rates. To address this, researchers have investigated the use of deep learning (DL) to identify pre-cancerous lesions specifically on the tongue.

Faced with a lack of existing information on tongue lesions within the oral cavity, the study utilised a specifically created dataset comprising portraits of patients' tongues. The team employed Convolutional Neural Networks (CNNs)—a class of deep learning algorithms designed to analyse visual imagery. They tested a wide array of architectures, including DenseNet, MobileNet, ResNet, and AlexNet, to determine which method offered the best diagnostic potential.

The evaluation focused on training and validation precision, alongside loss metrics. Among the various models tested using this customised data, the VGG16 architecture emerged as the top performer. It demonstrated a training accuracy of 97.66 per cent and a commendable validation accuracy of 89.06 per cent. These results suggest that applying algorithms like VGG16 to simple tongue imagery could provide a novel and effective method for the early diagnosis of oral cancer.