Research On Classification And Segmentation Algorithm Of Benign And Malignant Breast Tumors Based On Improved TransUNet

In 2020,the incidence and mortality rate of breast tumors in the world reached a record high.In traditional ultrasound examination,doctors mainly rely on professional knowledge and experience to judge tumor lesions,and there is strong subjectivity.Deep learning has become a research trend in auxiliary medical image analysis.In this paper,we intend to study a benign and malignant classification and segmentation algorithm for phacobreast tumors based on the improved Trans UNet.Firstly,according to the classification of the course of the disease,the experimental comparative analysis was carried out according to the classification scheme design,and the classification experimental results showed that the accuracy of the Efficient Net-V2 verification set was the highest,reaching 93.49%,and the model was the smallest,only 82.6M,so Efficient Net-V2 was selected as the benign and malignant classification model for phacobreast tumors.In order to obtain the specific breast tumor lesion area,this paper carried out the segmentation experiment of benign and malignant regions on the basis of classification.Secondly,for the task of segmentation of lesion area,six basic segmentation experiments are compared and analyzed,and the results show that Trans UNet has the highest accuracy in the verification set,with Io U(Intersection of Union)being 76.71%,Dice being 84.77%,and the model size being 150 M.However,there are still difficulties such as large models and poor edge segmentation.Aiming at the difficulties of the above model segmentation,this paper conducts experimental comparative analysis through a variety of backbone network design models,and the results show that the model with Efficient Net-V2 as the backbone network of TransUNet has an accuracy of 81.11% and a model size of 102 M,which realizes the purpose of lightweight.Experimental comparative analysis is carried out through a variety of attention mechanism design models,and the results show that the ACMix hybrid attention mechanism replaces the original single Transformer self-attention mechanism,and the accuracy reaches83.2%,which realizes the purpose of refining the edge segmentation accuracy.By using the CARAFE(Channel Attention Replenished by Adaptive Feature Enhancement)upsampling design instead of traditional sampling,and the ablation experiment results show that the final improved algorithm in this paper has an Io U of 84.74%,Dice of 90.7%,and a model size of105 M on the verification set,compared with the original Trans UNet segmentation algorithm.The Io U accuracy is improved by 8%,Dice is improved by about 6%,and the model is reduced by 45 M,which verifies the effectiveness of the algorithm improvement.In summary,the classification of the course of phacopromammary tumors and the related experimental analysis of lesion areas in this paper,and finally the improved algorithm can effectively carry out the classification and segmentation of phacobreast tumors as an auxiliary diagnosis,which has potential clinical application value.