Optimized deep learning networks for segmentation and classification of thyroid nodules in ultrasound images

In recent years, the incidence of thyroid nodules has been continuously increasing. With the increase in people's iodine intake and the growing environmental radioactive pollution, the risk of malignant transformation has also been increasing, posing a serious threat to human health. In the actual clinical diagnosis of thyroid nodules, doctors often use ultrasound imaging for detection. This study aims to accurately segment the thyroid nodule region and distinguish between benign and malignant nodules in ultrasound images using deep learning methods, thereby assisting doctors in improving diagnostic efficiency. To accurately segment the nodular region, this study utilizes the Mask R-CNN network to perform segmentation on ultrasound images and optimizes the backbone network to enhance its feature extraction capabilities. A branch is added to the feature pyramid model of the original backbone network, and all the output feature maps are then fused to obtain fused features, which are subsequently output. This achieves multi-scale feature fusion and balances the information differences among the output feature maps. To achieve accurate identification of benign and malignant thyroid nodules, this study applies the DenseNet network to detect ultrasound images and optimizes its network structure. Five branches are added to the original network to fuse the input and output feature maps, transmitting a large amount of low-level texture information. Attention mechanisms are also introduced between each Dense block and adjacent transition layers to retain effective image features and balance the weights of the fused feature maps, allowing the network to better utilize the global information from the feature maps. The improved network models are tested using nodule images, the improved Mask R-CNN model exhibits high accuracy. The improved DenseNet can provide reference for the clinical diagnosis of thyroid nodular diseases and enhance the diagnostic efficiency of doctors.