The Application of Texture Classification Network with Hybrid Adaptive Wavelet in Graves' Disease Ultrasound Diagnosis

Graves' disease is a common disease and ultrasonic examination is an effective means of its clinical diagnosis. One of the most common automatic diagnosis methods employed for Graves' disease is to classify ultrasound images using convolutional neural networks (CNNs) such as ResNet. Such techniques rely entirely on CNNs to automatically extract features, which could lead to the lack of an efficient application of texture features in ultrasound images. Wavelets, well known for their exceptional descriptive ability on texture features, could alleviate this limitation. In this paper, a CNN integrated with wavelet transform is applied to the diagnosis of Graves' disease. The model features a parallel wavelet branch within the CNN and includes a learnable Lifting Scheme module explicitly designed to extract wavelet features. This approach can simultaneously analyze texture features in both spatial and frequency domains, with implications for interpretable scientific reference in diagnosis. Experiments are conducted on a data set of size 214 and the designed network, with an accuracy of 97.270% and a recall rate of 95.603%, provides more effective diagnostic capability compared to ResNet, hence demonstrating the effectiveness of the proposed approach.