An improved CNN-based thyroid nodule screening algorithm in ultrasound images

Abstract

In recent years, the incidence of thyroid nodules has continuously increased. Ultrasound imaging, a preferred method for the clinical diagnosis of thyroid nodules, has the advantages of low cost, non-invasiveness, and real-time display. However, ultrasound imaging has certain issues, such as low contrast and difficulty in the early screening of nodules, which limited its effectiveness. To address these issues, we proposed a nodule detection method for thyroid ultrasound images that applies deep learning and feature extraction mechanisms. Taking YOLOX-M (Size-Medium) as an example, we made improvements to the original algorithm, making it representative of convolutional structure algorithms. To obtain good shallow features, we introduced Involution to increase the spatial specificity of features and provide rich nodule detail. To improve the early screening accuracy of nodules, deformable convolution was introduced in the backbone network for early detection of nodules with different geometric. To alleviate the effects of complex background information in thyroid ultrasound images on the detection of nodules, we introduced an attention mechanism before the feature fusion module to remove redundant information. The experimental results showed that the mAP (All nodules) and mAP@S (Early nodules) of the improved thyroid nodule detection algorithm reached 75.7% and 67.0%, respectively. Compared to the original YOLOX-M algorithm, the detection accuracy was improved by 3.5% and 1.8%. Finally, we integrated this method into other convolutional structure algorithms, which had different degrees of improvement and performance in nodule screening.