Background: Machine learning is an emerging technology in health care field with aim of fundamentally revamping the traditional system and aiding medical practitioners. The histopathological analysis of oral cancers is crucial for pathologist to ascertain its grading. Therefore, this study attempts to grade the various stained tissue samples of OSCC (Oral Squamous Cell Carcinoma) patients using different deep-learning models.
Methods: A dataset of 120 histopathological images of OSCC was collected and classified as well-differentiated (40), moderately differentiated (40), and poorly differentiated (40) according to Broders' grading system. The CNN (Convoluted neural networks) architectures were based on the pre-trained neural network models VGG16 (Visual Geometry Group16), VGG19 (Visual Geometry Group19), RESNET50 (Residual Network50), and DENSENET121 (Dense Network121) models for image analysis.
Results: At a magnification of 4x, all four models achieved the highest test accuracy of 66.67%. DENSENET121 scored the highest validation accuracy of 81%. At 10x, RESNET50, VGG19, and DENSENET121 achieved the highest test accuracy of 90.9% and VGG19 achieved the highest validation accuracy of 83.3%. At 40x, the highest test accuracy of 70% was achieved by RESNET50 and DENSENET121. The validation accuracy was comparable between RESNET50, VGG16, and VGG19.
Conclusion: The grading of tissues with the help of deep learning in digital imaging and computational aid in the diagnosis can help in timely and effective prognosis and multi-modal treatment protocols for oral cancer patients, thus reducing the operational workload of pathologists. By systematically evaluating model performance and addressing concerns about overfitting, we develop robust models suitable for medical diagnosis.
Keywords: CNN; OSCC; histological analysis; machine learning; transfer learning.
Copyright © 2024. Published by Elsevier Masson SAS.