Enhancing thyroid nodule assessment with deep learning and ultrasound imaging

Abstract

The thyroid is a tiny, butterfly-shaped gland in the neck which produces hormones that are essential for controlling the body's various metabolic processes. Thyroid nodules, which are abnormal growths or lumps in the thyroid gland, are common thyroid illnesses, as are hypothyroidism, hyperthyroidism, and both. Thyroid issues are most commonly identified and categorised using thyroid ultrasonography (USG) images. They can have a range of effects on the body's metabolism and overall health. Developments in artificial intelligence (AI), particularly deep learning (DL), are helping to identify and measure patterns in clinical images because of DL's capacity towards pull out hierarchical attribute representations from images without the need for annotated images. Minimizing unnecessary fine needle aspiration (FNA) requires the essential identification of as many malignant thyroid nodules as possible, distinguishing them from benign ones. This research work introduces a technique for thyroid nodule identification in USGs, employing DL to extract relevant features. Three pre-trained DL models, namely ResNet-18, VGG-19 and AlexNet were fine-tuned before using for classification of thyroid USG images. The models' testing and training were done with Digital Database of Thyroid Ultrasound Images (DDTI) which is gold standard dataset. The results demonstrate a classification accuracy of 97.13%, 90.31% and 83.59% with ResNet-18, VGG-19 and AlexNet, respectively. The experimental findings affirm that the pre-trained network model ResNet-18 achieves superior classification performance compared to VGG-19 and AlexNet.