An improved deep convolutional neural network fororal cancer detection using pretrained models

Abstract Purpose: Oral cancer is a type of cancer that arises from Oral Squamous Cell Carcinoma (OSCC) that develops in the mouth. Oral cancer is a major global public health issue, which emphasizes the urgent need in the research domain for targeted and effective approaches that enable timely identification. The current diagnostic approaches has a limitation of delay and inefficiency. Artificial Intelligence (AI) is nowadays extensively used the cancer diagnosis and prognosis can provide a fast results that helps in early detection of cancer and improves the survival rate of the cancer patients. Deep learning techniques offers several pretrained models in automating the cancer detection. The research focus on developing a light weight architecture with improved results and low computational costs based on DenseNet architecture of the Convolutional Neural Network (CNN). Methods: The number of dense blocks are reduced without compromising the classification of histopathology images as benign and malignant. The histopathology images are preprocessed in a way that is suitable to fed to a neural network. The hyper parameters of the model is fine tuned to obtained the better results. Results: The proposed model is evaluated with the standard performance metrics and compared with the other pretrained models. The proposed model provides 98.96% of classification accuracy in training and 82.49% in validation. The loss also has been reduced with a good precision of 0.98, recall of 0.76 and F1 score of 0.82. Conclusion: The proposed model also takes care of overfitting and vanishing gradient problem that could hamper the models performance. This will also help a medical practitioner can utilize these findings to formulate initial decisions and plan treatment for oral cancer.