Research Of Medical Image Segmentation Algorithms Based On Feature Recombination And Context Information Enhancement

With the development of science and technology,computer-based digital image processing technology has become increasingly mature and has gradually become an indispensable tool in all walks of life.Especially in the medical field,auxiliary diagnosis based on medical image processing technology has greatly improved the efficiency and accuracy of disease examinations,providing important reference value for clinicians to formulate reasonable medical plans.Medical image segmentation is an important part of auxiliary diagnosis.It is a technology that segments specific target areas from two-dimensional or three-dimensional medical images and quantitatively analyzes the structural characteristics of the target tissue.In particular,automated segmentation can not only reduce the labor burden and time consumption of doctors manually delineating target areas,but also provide convenience and reference for preoperative analysis,intraoperative guidance,and postoperative evaluation.In recent years,data-driven learning methods have received widespread attention and discussion due to their excellent image feature extraction capabilities.However,in some medical fields,it is often difficult to obtain data with rich annotation information,resulting in low generalization capabilities of models that use small sample data to drive learning.How to better extract rich medical image features and improve the robustness and anti-interference of the model is what everyone is pursuing;on the other hand,with the development of imaging technologies such as CT,the volume of cancer images is getting larger and larger.,brings opportunities to the development of data-driven learning methods,and at the same time,also brings challenges to automated liver tumor segmentation methods driven by CT and other imaging data.From an imaging point of view,image data is easily affected by factors such as impulse noise and differences in imaging protocols,resulting in a large difference in available image quality.From the perspective of the target itself,the density is heterogeneous,the boundary area is blurry,and the boundary contour is inconsistent.Reasons such as rules and scale diversification have led to a decrease in the accuracy of automated segmentation models.In particular,the ubiquity of small target tumors has intensified the difficulty of segmentation.In order to deal with the problems faced in the segmentation process,we have done the following research work:(1)Based on the problems of weak generalization ability of learning models trained on small sample data and insufficiently rich captured image features,we proposed a novel feature reorganization strategy.The idea is to construct multiscale features at different levels of the target,break up and reorganize the features at the same scale through feature reorganization,and use multi-level information at different scales to enrich the feature space of the image,thereby improving the accuracy of model segmentation.Stickiness and accuracy.The model design based on this strategy achieves information interaction between features as much as possible.The result not only has good performance in two-class segmentation,but also has good performance in multi-class segmentation.(2)Liver tumor CT image data has different image quality differences due to different imaging equipment and protocols:which adds trouble to the training of segmentation models.Secondly,the tumor target area to be segmented is not homogeneous,and the boundaries are fuzzy,irregular,etc.factors further exacerbate the difficulty of segmentation.In order to solve the problem of large differences in image quality,this article established a set of original data processing methods to unify the data distribution between different devices through truncation,normalization,and de-energization.Secondly,for the difficulties in liver tumor segmentation,this article uses The multi-scale perception module reduces the loss of target detail information and enriches semantic features at different scales.At the same time,the dual-split attention module is used to embed the position information of the target when calibrating the channel response of the feature map.These strategies enhance the context of the target.information,improving segmentation capabilities.In clinical applications,the statistics of the number of tumors can be achieved through three-dimensional reconstruction of the segmentation results,which provides a reference for tumor diagnosis.(3)The liver tumor segmentation model based on multi-scale perception and bi-split attention has been greatly improved compared to the contrast method.However,in the evolution of tumors,the characteristic information of small-scale tumors is often ignored or filtered,which is not conducive to early stage Detection of tumors and situations where large and small tumors vary greatly in size.Combining these problems and taking into account the local characteristics of the tumor itself,especially small tumors,this paper proposes a local similarity embedding strategy to establish the dependence between local and global lesions and better deal with the difficult problem of segmentation of small tumors.In addition,the multi-core pooling module and interactive additional attention also enhance the feature extraction and purification capabilities of the model,enriching the contextual information of the tumor target,thereby further improving the segmentation accuracy.The model extracted in this article has not only been verified and compared in different public data sets,but also led the generalization test of the model on some actual clinical data sets.The results have demonstrated the superiority and robustness of the proposed method.