Automatic Thyroid Nodule Detection in Ultrasound Imaging With Improved YOLOv5 Neural Network

Abstract

With the incidence of thyroid cancer increasing dramatically, the burden of sonographic diagnosis lies heavy for radiologists. An automatic computer-aided diagnosis system with both precision and efficiency is in demand. This retrospective study included 191 ultrasound images of 171 patients (85 benign and 86 malignant) in Wenzhou Central Hospital. An improved You Only Look Once version 5 neural network (improved YOLOv5) is proposed in this work. It comprises the coordinate attention (CA) module and the label smoothing regularization (LSR) module, in which the CA module enables the network ability of positional information extraction. The improved neural network correctly recognizes the lesion area and nodule type with a mean average precision (mAP) of 95.3% in 8.4 ms on the test set. The ablation experiment demonstrates that the integration of the CA and the LSR module cost 1.3 ms extra inference time per image in exchange for raising the mAP by 4.4%. Afterward, 10 ultrasound images with wrong nodule types are added to the dataset for training, the result shows that the LSR module can significantly prevent the network from being misled by the bug data. Compared with other state-of-the-art networks, the improved network has superiority in both precision and robustness for diagnosing benign/malignant thyroid nodules with only a small dataset. The proposed network also has the potential to transfer to other sonographic diagnosis tasks.

Keywords

• Computer-aided diagnosis
• thyroid nodules
• ultrasound
• YOLOv5 network
• attention module
• label smoothing