A Convolutional Neural Network Based Deep Learning Algorithm for Identification of Oral Precancerous and Cancerous Lesion and Differentiation from Normal Mucosa: A Retrospective Study

Oral cancer is the sixth most common cancer associated with high diseases related mortality. The prime reason is that more than two thirds of patients are diagnosed at later stages of cancer. Majority of oral cancers are preceded by noticeable changes in the oral mucosa known as oral potentially malignant disorders (OPMDs). Early diagnosis of OPMDs can elude cancer development in 88% of cases. Artificial intelligence (AI) has gained popularity in the field of medicine including oncology and has shown efficacy in the diagnosis and prediction of cancer prognosis. In the present study, pre-trained convolutional neural networks (CNNs) are used for identifying oral pre-cancerous and cancerous lesions and to differentiate them from normal mucosa using a dataset of clinically annotated photographic images. This study was conducted on clinical photographs of patients who presented with oral squamous cell carcinoma and OPMDs. A comparative analysis of these photographs were done with photographs of normal oral mucosa. Transfer learning using various pre-trained CNN architectures was employed for image classification. An accuracy of 76% for VGG19, 72% for VGG16, 72% with MobileNet and 68% for InceptionV3 and 36% with ResNet50 was obtained. In the present study, VGG19 exhibited good performance when compared to other models.