Automated Detection of Oral Squamous Cell Carcinoma using Transfer Learning Models from Histopathological Images
Gurjot Kaur, Neha Sharma · 2024
This work aims to diagnose Oral Squamous Cell Carcinoma (OSCC) from histopathology images using transfer learning models especially, EfficientNetB3 and DenseNet201. There are 5,192 histological pictures in the dataset. From the dataset, the DenseNet201 model split a 120-image test set, a 126-image validation set, and a 4,946-image training set. The EfficientNetB3 model used similarly 3,634 images for the training set, 779 images for the validation phase set, and 779 images for the test set. With batch sizes of 32 and 128 respectively, these models were trained across 50 epochs using the Adam optimizer to adjust the model parameters. With a test accuracy of 0.9127, a validation accuracy of 0.99083, and an Area Under the Curve (AUC) score of 0.866, the DenseNet201 model showed notable ability in differentiating between malignant and non-malignant tissue samples, therefore suggesting a high degree of precision. With a training loss of $\mathbf{0 . 5 0 4 2}$, a training accuracy of $\mathbf{0 . 9 4 5 1}$, a validation loss of 0.5030, a validation accuracy of 0.9474, a test loss of 0.6893, and a test accuracy of 0.9217 the EfficientNetB3 model displayed amazing performance. EfficientNetB3's outstanding test accuracy emphasizes its possible use as a good model for histological image interpretation. These results show how well transfer learning models especially, DenseNet201 and EfficientNetB3 might improve OSCC detection accuracy and efficiency from histological images. These models' strong performance points to their major contribution to creating automated diagnosis tools in oncology, enabling early detection and treatment planning for OSCC. By introducing such cutting-edge computing methods into clinical practice, pathologists' diagnostic load may be reduced, diagnosis consistency could be raised, and finally, better patient outcomes could result. Especially in cancer identification and analysis, this study marks major progress in the use of artificial intelligence in medical diagnostics.