Non-invasive Primary Screening of Oral Lesions into Binary and Multi Class using Convolutional Neural Network, Stratified K-fold Validation and Transfer Learning

Abstract

Objectives: To develop a deep learning method using camera images that can effectively detect the preliminary phase of oral cancer, which has a high rate of morbidity and mortality and is a significant public health concern. If left untreated, it can result in severe deformities and negatively affect the patient's quality of life, both physically and mentally. Early detection is crucial owing to the rapid spread of the disease, where biopsy is the only option left. Therefore, it is essential to identify malignancies swiftly to prevent disease progression non-invasively. Methods: Three different scenarios are used in this study to analyze samples: CNN architecture, stratified K-fold validation, and transfer learning. For automated disease identification on binary datasets (normal vs. ulcer) and multiclass datasets (normal vs. ulcer vs. Leukoplakia), camera images are pre-processed with data augmentation. As a feature extractor in the model, transfer learning is used with pre-defined networks such as VGG19, InceptionNET, EfficientNET, and MobileNET weights. Findings: Using the proposed CNN architecture, the F1 score for image classification was 78% and 74% for photos showing hygienic mouths or ulcers, and 83%, 87%, and 84% for images showing normal mouths, ulcers, and leukoplakia. Using stratified 3-fold validation, the results were improved to 97%, and an EfficientNET achieves the highest results in a binary F1 score of 98% and a classification with multiple classes F1 scores of 98%, 87%, and 91%, respectively. Novelty: Previous studies have mostly concentrated on differentiating oral potentially malignant diseases (OPMD) from oral squamous cell carcinoma (OSCC) or on discriminating between cancerous and non-cancerous tissues. The objective is to diagnose patients with non-invasive procedures to classify ulcers, healthy mouths, or precancerous type "Leukoplakia" without requiring them to visit a doctor. Keywords: CNN, Transfer Learning, Oral Cancer, Ulcer, Leukoplakia, Stratified K-­fold validation

A Review on Non-invasive Tools and Deep Learning/Machine Learning Methods for the Early Identification of Oral Cancer

“…The same researchers looked at pictures to find diseases automatically on binary and multiclass datasets using CNN architecture, stratified K-fold validation, and transfer learning in a different study.The proposed CNN architecture achieves F1 scores of 84%, 78%, and 87% for hygienic mouths or ulcers, and 83%, 87%, and 84% for normal mouths, ulcers, and leukoplakia, respectively. The idea is to classify ulcers, healthy mouths, and precancerous type "Leukoplakia" using non-invasive techniques, hence diagnosing patients without the need for them to see a physician [22]. Roshan Alex and his colleagues then went one step further, applying the Bounding Box Method to composite annotations created by combining annotations from multiple clinicians and further applying transfer learning to multiple images, yielding encouraging results [23].…”