Research On Transformer-based Semantic Segmentation Algorithm For Medical Images

A precise segmentation of lesions is essential for the quantitative analysis of many diseases.Early detection of lesions can significantly improve the survival rate of patients.With the rise of artificial intelligence technology in recent years,researchers have attempted to combine semantic segmentation with medical image analysis,aiming to improve the efficiency of doctors in segmenting lesions during clinical examinations.As breast cancer is one of the most diagnosed diseases,it is important to find solutions for the difficulties in diagnosing breast cancer lesions.This paper focuses on the localization of microscopic lesions and the ambiguous boundary problem caused by the growth and spread of tumors in the medical imaging diagnosis of breast cancer.To address the above two key problems in breast cancer diagnosis,two Transformerbased breast cancer lesion segmentation models are proposed from different directions in this paper.As a summary,the main research findings of this paper can be summarized as follows:(1)Magnetic resonance imaging is an important imaging method for breast cancer detection,but the small size of breast cancer lesions in magnetic resonance images makes them difficult to locate.In this paper,an axial attention-based segmentation model is proposed to solve the problem of segmenting and localizing breast cancer tumors in magnetic resonance images.In contrast to other medical segmentation methods,this paper uses only the attention mechanism to construct the segmentation model,and explores the potential of the Transformerbased model for breast cancer segmentation.For large scale medical images,this paper decomposes the traditional attention mechanism on two-dimensional images into two onedimensional attentions along the image height and width axes,thus reducing Transformer’s computational complexity on large sized medical images.Additionally,axial attention calculations are improved using relative position information and gate units to limit inaccurate position information and highlight lesions in patches.According to tests on a breast cancer MRI dataset,the axial attention-based Transformer segmentation algorithm performed well in localizing microscopic lesions in breast cancer.(2)In clinical diagnosis,the accurate localization of tumors alone cannot achieve the goal of assisting radiologists to improve the efficiency of diagnosis;due to cancer cells’ inherent spreading characteristics,the boundaries between lesions and normal tissue are ambiguous.As a solution to this challenge,a boundary-aware Transformer segmentation model is proposed.In this paper,a Transformer segmentation model based on edge attention is proposed to improve the segmentation accuracy of tumor fuzzy boundaries by adding edge key points and constructing a loss function for edge segmentation.Further,by adding reference points to the decoder part of the model,accurate segmentation of lesion edges and accurate localization can be achieved.According to the experimental results on three sets of breast cancer magnetic resonance images and breast cancer ultrasound images,the proposed model has good segmentation performance and effectively solves the problem of ambiguous boundaries between breast cancer lesions in different images,thereby ensuring accuracy and integrity in lesion segmentation.