Ultrasonic Thyroid Nodule Benign-Malignant Classification with Multi-level Features Fusions

A novel classification model that integrates shallow and deep features is proposed. Firstly, traditional image processing methods, such as LBP and Canny operator, are utilized to extract shape, texture, and edge features from the region of interest. Then, a novel FCST model composed of a residual convolution branch and a Swin Transformer branch is designed for extraction deep features of ultrasound nodule images. The residual convolution branch extracts local features and passes them layer by layer to the Swin Transformer, which compensates for the missing information of local details and enables the model to capture both local details and global dependencies. Finally, the shallow and deep features are fused for classification to obtain the identification results of benign and malignant nodules. Experimental results demonstrate that the fusion of deep and shallow features can effectively enhance the network's capability of identifying nodules, significantly improving the accuracy of identifying benign and malignant nodules.