Classification of Cervical Intraepithelial Neoplasia (CIN) using fine-tuned Convolutional Neural Networks
Oluwatomisin E. Aina, Steve A. Adeshina, Adeyinka P. Adedigba, Abiodun Musa Aibinu
Abstract
Convolutional Neural Network (CNN) is considered one of the most successful deep learning techniques used in classification or diagnosis of medical images. However, CNN requires a high computational resource and time; and a large dataset which most medical images (cervix) do not possess. In order to compensate for these shortcomings, we propose an optimized fine-tuned CNN model to classify cervix images into Cervical Intraepithelial Neoplasia grades (CIN 1,2,3) normal and cancerous cervix images. This classification ensures that patients are diagnosed correctly, and appropriate treatments are administered. Deep learning techniques such as Data Augmentation, 1 cycle policy for optimal learning rates selection, Discriminative Fine-Tuning, Mixed Precision Training were used to optimize the fine-tuned DenseNet CNN model. The model achieved 96.3% accuracy, the specificity of 98.86%, and sensitivity of 94.97% on the datasets.