Fully Automated Thyroid Nodule Detection, Segmentation, and Classification in Ultrasound Images Using Deep Learning and Image Processing

Abstract

Thyroid cancer has a high prevalence all over the world. Accurate thyroid nodule detection and diagnosis in early stages leads to effective treatment and decreases the mortality rate. However, thyroid nodule detection and assessment using ultrasound imaging is a very challenging task, even for experienced radiologists, due to the ultrasound image characteristics and variations in thyroid nodule sizes and appearances. Existing Computer-Aided Diagnosis (CAD) systems are not fully automated and also have limited performances. This thesis presents a fully automated thyroid CAD system to assist radiologists. The proposed CAD system consists of four components: nodule detection, nodule segmentation, thyroid segmentation, and nodule classification from thyroid ultrasound scans acquired through ultrasound examination of the thyroid. For nodule detection, a novel one-stage detection network, TUN-Det, is proposed, which introduces Residual U-blocks (RSU) to built the TUN-Det backbone, and presents a newly designed multi-head architecture comprised of three parallel RSU variants to replace the plain convolution layers of both the classification and regression heads. Residual blocks enable each stage of the backbone to extract both local and global features, and the multi-head design embeds the ensemble strategy into one end-to-end module to improve the accuracy and robustness by fusing multiple outputs generated by diversified sub-modules. TUN-Det achieves very competitive results against the state-of-the-art models on the overall Average Precision (AP) metric and outperforms them in terms of AP35 and AP50. For nodule segmentation, a residual dilated U-Net, resDUnet, is proposed, which has a residual structure, and also dilated convolution layers are embedded in the bottleneck part of the network. Residual connections lead to consistent training and dilated convolution layers generate richer multi-scale features. Our resDUnet achieves a high Dice score and much smooth visual results. For thyroid gland segmentation in ultrasound sweeps, LSTM-UNet is proposed, which uses time-distributed convolution blocks and bidirectional convolutional LSTM in the U-Net. The building blocks extract spatial-temporal information and consider the inter-frame correlation of consecutive frames. LSTM-Unet avoids the under-segmentation problem, which is a common issue in thyroid segmentation methods. For the nodule classification component, two rule-based classifiers are proposed for nodule composition and nodule margin, which use different image processing techniques and decide based on the pre-defined rules. The rules are defined based on the clinical definitions. All Experimental results indicate the promising performance of the proposed CAD system in clinical applications.