An Improved Deep Convolutional Neural Network Model for Thyroid Nodule Classification

Recently, Convolutional Neural Networks (CNNs) have become a methodology of choice for processing of medical images as they have a number of advantages compared to traditional machine learning techniques. In this paper, a deep Convolutional Neural Network for Thyroid Nodule Classification, CNN-TNC, is proposed using three different architectures to classify Thyroid Nodules using Thyroid Ultrasound (TUS) Images. Different data augmentation techniques like affine transformation that includes rotation, translation, scaling and shearing effects are carried out to increase TUS image dataset. At first, CNN-TNC is developed using custom architecture with 13-layer network which gave training accuracy of 79.65% and test accuracy of 72.50%. Further, to improve diagnostic accuracy of CNN-TNC, Transfer Learning (TL) technique is adopted using ResNet-18 and VGG-16 architectures. CNN-TNC implemented using ResNet-18 employing TL yielded train and test accuracy of 98.60% and 98.64% respectively. Also, CNN-TNC employing TL with VGG-16 resulted in an accuracy of 97.55% and 97% for train and test data respectively. CNN-TNC implemented using ResNet-18 and VGG-16 employing TL technique found to be the best when compared to CNN-TNC implemented with custom architecture. Obtained results demonstrate how fine-tuning pretrained CNN models in a layer-wise method leads to incremental performance improvement when training data is limited.