Feature Extraction using Hybridized Transfer Learning Approach for Oral Cancer Diagnosis

Oral cancer diagnosis at earlier stage is very crucial to decide the treatment procedure and to avoid mortality due to this malignant disease. Histopathological imaging is one among modalities widely used by the clinicians to identify the oral cancer. This research focuses on utilizing the hybridized version of transfer learning to extract the deep features from the histopathological oral images for oral cancer detection. Three transfer learning approaches namely VGG16, MobileNetV2, and ResNet50 are used to extract the features individually and then variance based filter technique is used to choose the appropriate features. The experiment is done using 439 oral cancer images and 89 normal images which are obtained from an online repository. The suggested approach offers ${9 7 \%}$ accuracy while the accuracy attained by stand-alone transfer learning models are less than 90%.