An ultrasonic image lesion segmentation method based on CNN-Transformer hybrid network feature transformation

Abstract

The application discloses an ultrasound image lesion segmentation method based on a CNN-Transformer hybrid network feature transformation, which comprises the following steps: after pre-processing the input ultrasound image, inputting the hybrid network with an encoder-decoder structure; extracting local features and global semantic features through a parallel double-encoder module, and realizing adaptive conversion between the two branches through a feature conversion module; purifying the global semantic features through a feature enhancement and conversion module and converting the global semantic features into adaptive features; performing feature upsampling and fusion through a convolutional neural network decoder, and outputting a segmentation probability map; and obtaining a lesion region mask after binarization. The method can accurately capture lesion features, improve the accuracy and efficiency of ultrasound image lesion segmentation, has good generalization and robustness, and can still maintain high segmentation accuracy on an external verification set and a data set containing normal image interference; the balance between segmentation accuracy and calculation efficiency is realized under a reasonable parameter scale, and the method is suitable for the practical application of clinical ultrasound images.

Images