Optimizing CNN based model for thyroid nodule classification using data augmentation, segmentation and boundary detection techniques

Thyroid nodule is an asymptomatic disorder which mostly occurs due to high production of thyroid hormones from the thyroid gland. The diagnosis is usually made by the radiologist and endocrinologists which heavily relies on their experience and expertise. Ultrasonography is one of the principal means for the initial assessment of nodules which is mainly performed when there is suspect of formation of nodules. In this research work, an optimized convolutional neural network model is proposed for the identification of thyroid nodules using various deep learning techniques like dense neural network, Alexnet, Resnet-50 and Visual geometry group-16. A total of 295 public and 654 collected thyroid ultrasonography datasets are considered in this work. The proposed model is evaluated on 1475 public and 3270 collected thyroid ultrasonography datasets with data augmentation technique. We experimentally determined the best optimized value for learning rate and drop out factor to enhance the performance of the models. The proposed model has achieved an accuracy of 93.75%, sensitivity of 94.62%, specificity of 92.53% and f-measure of 94.09% on the public dataset in experiment-I and an accuracy of 96.89%, sensitivity of 97.80%, specificity of 94.73% and f-measure of 97.26% on the collected dataset in experiment-II. The proposed model has shown an improvement of (4.57%, 7.84%), (5.06%, 8.24%), (4.43%, 6.63%) and (4.66%, 7.83%) in terms of accuracy, sensitivity, specificity and f-measure on (dataset −1, dataset-2) against other state of the art models.