Oral Cancer Detection Using Convolutional Neural Network(CNN)

Abstract

Oral squamous cell carcinoma (OSCC), the most widespread type of oral cancer, necessitates prompt biopsy evaluation in order to increase patient survival rates. However, healthcare professionals frequently disagree on subtle cell changes when performing standard histological (H&E)-stained slide evaluations. To overcome this issue, we developed a convolutional neural network (CNN) classifier using EfficientNetB3 pre-trained weights to classify oral samples as either normal or cancerous with high reliability. We created a composite image collection by combining publicly available images into a 5,192-image dataset that was randomly divided into 70%/15%/15% (3,634 training/779 validation/779 test) images, cropped to 224×224 pixels, normalized, and augmented (flipped) to create a robust training set. Our classifier was built upon the core of EfficientNetB3, adding a batch normalization, and using Adamax optimization for 100 epochs. The final performance of the classifier on unseen test image data was 98.33% accuracy and a balanced precision/recall/F1 score of 98% for each class (as shown by the confusion matrix) indicating the classifier is ready for clinical use in providing consistent second opinion results to help extend the capacity of pathologists in heavily burdened hospitals in India.