Thyroid nodule classification from ultrasound images using deep learning

As technology continues to evolve, the application areas of artificial intelligence are expanding, particularly in the healthcare sector. Deep learning models applied to medical imaging are increasingly being adopted for disease detection and interpretation. This trend is particularly evident in numerous studies focusing on the detection of thyroid nodules, whose incidence has been rising in recent years. Ultrasound, a common diagnostic method for detecting and classifying thyroid nodules, involves a subjective and time-consuming evaluation process based on expert experience. Therefore, to improve the traditional diagnosis process and reduce unnecessary biopsies following false positive diagnoses, a deep learning approach has been developed. In this study, we worked with a dataset containing 445 thyroid ultrasound images from 110 patients in Türkiye, including 86 benign and 24 malignant cases. The model training, validation, and testing were conducted using three separate datasets containing 305, 40, and 100 images, respectively. To address the class imbalance, data augmentation was applied, and external malignant data from the National University of Columbia were added. A Convolutional Neural Network architecture based on deep learning was used to classify thyroid nodules. To solve the problem of insufficient data, transfer learning methods using ImageNet and then RadImageNet, which only contains medical images, were compared. Various models such as EfficientNetB0, ResNet50, DenseNet121, and Xception were evaluated. Following hyperparameter optimization, the ResNet50 model achieved the best results with an accuracy of 91\%, a precision of 87\%, and a sensitivity of 65\%. Images in the test set have been classified as malignant or benign by expert endocrinologists and radiologist. It has been determined that our deep learning-based model provides a higher accuracy rate.