Feature Recognition of Ultrasound Breast Images Based on Improved DSA-U++
Weichun Guan, Yuanyuan Zhang, Xiaosen Lv, Shuyun Lv, Mao Lin, Zhengyang Liu · Frontiers in Computing and Intelligent Systems · 2025
The aim of this project is to recognize features in ultrasound breast images. And an improved DSA-U++ model is proposed based on the image classification task, the traditional U-Net++ in the encoder part in the face of ultrasound images there is a lack of feature extraction, to solve this problem we use Resnet50 as the backbone of the extraction, in order to further enhance the ability of the feature learning, we also introduced the ASPP module, to help capture contextual information at different scales, and a module R-AS is designed to enhance the model multi-scale perception ability. information to enhance the model's multi-scale perception ability and a module R-AS. the output of R-AS after five stages is used as the encoding part of U-Net++, and the feature information extracted by the encoder is reconstructed and enhanced in the decoder part. In order to reduce the computational complexity and the number of parameters, and to maintain a certain feature extraction ability, we replace the traditional convolution in the decoder part of U-Net++ with a depth-separable convolution, which is experimentally validated on the AI Algorithm Elite Challenge dataset, which consists of ultrasound breast images with four features, namely, orientation, edges, calcifications, and shapes, and is trained with the DSA-U++ model. After the DSA-U++ model is trained, it improves on several indexes, not only improving the recognition ability of subtle features in ultrasound breast images, but also effectively improving the performance of image classification.