Developing a Deep Learning Ultrasonography Model to Classify Thyroid Nodules as Benign

1.<Introduction>Thyroid cancer, or the occurrence of rapid cell growth in the thyroid gland located near the neck, is the fastest growing cancer among women. Papillary thyroid cancer leads to hormonal imbalances thus causing periods of fatigue, difficulty breathing, and an overall decrease in ones quality of life. <Objective>Unsurprisingly, the need for a quick diagnosis of thyroid cancer has become ever more important. Deep learning is a subset of machine learning that may improve the diagnostic performance of ultrasound, including in the evaluation of thyroid nodules (irregular growths that develop on the gland) for benign vs. malignant disease. Currently, the standard method to differentiate benign or malignant thyroid nodules is an invasive test known as Fine Needle Aspirations (FNAs) that is uncomfortable and may lead to unwanted outcomes such as bleeding, and infection. <Methods>We propose using a convolutional neural network with a foundation rooted in two pre-trained networks, VGG-16 and InceptionV3, to train a model on the publicly available Thyroid Digital Image Database to accurately classify thyroid nodules on ultrasonography into those that are malignant or benign; a non-invasive strategy. <Results>The data was augmented using industry-standard methods and provisioned into train/test/validation sets obtaining 88% accuracy and an area under the receiver operating characteristic (ROC) of 0.929 (signifying a high sensitivity). <Discussion>Based on the models results, it is conceived that a convolutional neural network can serve as an accurate classifier of malignant/benign thyroid nodules from an ultrasound and enhance a diagnostic strategy to reduce unnecessary FNA procedures.