Detection and classification of thyroid diseases using ultrasound images through deep learning techniques

Abstract

Thyroid diseases remain one of the significant public health issues around the world, with thyroid nodules estimated to be prevalent among approximately 30 to 50% of the adult population through ultrasound imaging screening processes. Determination of the nature of the thyroid nodules in a patient's body is vital for ensuring effective treatment options since early detection of any malignant nodule will result in positive treatment results. The problem with such an analysis is the fact that the visual appearance of both benign and malignant nodules resembles each other in ultrasonic images. Therefore, a two-stage deep learning-based model for automatic detection and classification of thyroid nodules in ultrasound images was developed in this paper. At the first stage, the proposed framework applied Attention U-Net segmentation to identify the region of interest (ROI) for thyroid nodules in ultrasound images. Expert-drawn pixel-level annotations of thyroid nodules were not available; hence, pseudo-labels were used in training. They were obtained by applying CLAHE, Otsu thresholding, and morphological operations. In the second stage, a customized CNN network was applied to classify the segmented thyroid region into either benign or malignant classes. Data augmentation using the Mixup and CutMix techniques helped reduce the risk of overfitting, enhancing the model's generalizability. This experiment was tested on 480 ultrasound images gathered from two teaching hospitals in Nepal. The proposed framework shows that segmentation with weak supervision along with classification using deep learning can provide efficient diagnosis of thyroid nodules even with minimal medical data annotation. Through 3-fold stratified cross-validation, the Attention U-Net achieved a Dice Score of 0.8772 and IoU of 0.7873 for segmentation, while the classification stage obtained an overall accuracy of 96.73%, a specificity rate of 97.89%, and a mean ROC-AUC of 94.60%.