Deep Learning-Based Ultrasound Image Detection and Recognition Method for Thyroid Nodules

Thyroid nodules are extremely prevalent and have a tendency to become malignant with age. The clinical diagnosis of nodules is influenced by subjective and objective factors such as medical conditions, image quality, and clinical experience. The rapid development of artificial intelligence in recent years has prompted deep learning algorithms to make great progress in computer-aided diagnosis, which can effectively improve diagnostic efficiency. The current development of convolutional neural network model tends to be a deeper network and more number of model parameters, this paper proposes a general framework that combines the SSD algorithm and the improved Inception-ResNet-V2 structure, which uses a smaller number of channels in the network and greatly reduces the number of parameters. In addition, the CBAM module was introduced to focus more on the lesion area and filter out the interference of background noise. In this paper, the desensitized thyroid ultrasound images in the hospital are used as the original data, the dataset is expanded by data enhancement methods on the original data, and the model is trained based on this dataset. To validate the model's optimal performance in deep networks, experiments compared the detection accuracy with different backbones. The experimental results showed that the mAP, accuracy, precision, and F1 score of the model proposed in this paper reached 93.1%, 90.9%, 95.7%, and 91.7% for the model at an IoU of 0.5, which can effectively assist the diagnosis of nodules.