Automatic Medical Image Segmentation Algorithm Based On Attention Mechanism

Medical imaging is a very important information carrier,which can help clinicians understand the patient’s condition and assist clinicians in planning treatment plans.Medical image segmentation technology also plays a vital role in clinical treatment,such as lesion size measurement,organ and lesion positioning,radiotherapy planning,auxiliary surgery,and anatomical structure research.At present,image segmentation technology based on traditional image processing and deep learning is widely used in medical image automatic segmentation.However,accurate medical image segmentation still has the following challenges.(1)The tissues and lesion structure of patients present a high degree of diversity and variability.(2)There are many imaging modes and imaging parameters for medical images,and different imaging modes will result in huge differences in medical images.(3)The imaging quality of medical images is not uniform,and there are many kinds of artifacts and noises that cause image blurring and other problems.Therefore,in view of the above-mentioned problems,researching corresponding solutions and developing accurate and effective automatic medical image segmentation algorithms have important research and application value,which is also the research topic of this article.This article takes medical images as the research object,and conducts research on brain tumor segmentation,skin lesion segmentation,three-dimensional vertebral body segmentation,organs-at-risks segmentation and cell nucleus segmentation.This article focuses on constructing the attention methods with interpretability and strong feature learning ability,and proposes three medical image segmentation algorithms based on deep learning.(1)In order to build a more effective high-and low-level feature fusion mechanism and enhance the coding quality and semantic consistency of high-and low-level features,this paper proposes an image segmentation algorithm based on cascaded high-and low-level feature enhancement.At the same time,a low-level feature enhancement module is proposed,which was used to embed high-level semantic information into low-level features;a high-level feature enhancement module is proposed,which is used to aggregate the optimal high-and low-level hybrid features and embed more effective spatial information into the high-level features.In addition,through visual analysis,the feature learning mechanism behind deep learning is explored,and efforts are made for the interpretability analysis of deep learning in the field of medical image processing.(2)In order to use the spatial similarity between attention modules to construct a unified and joint attention optimization architecture,this paper proposes an image segmentation algorithm based on cross-layer attention fusion.This method takes the cross-layer attention fusion architecture as the main body,which sequentially integrates the spatial attention weight maps of the attention modules,so that all attention modules in the network form a whole,and they learn and optimize together.In addition,a new Top-K exponential logarithmic Dice loss function is proposed,which is used to balance the network’s segmentation between samples of different sizes and samples of different difficulties.(3)In order to simplify the parameter selection of the network architecture and optimize the architecture of the attention module,this paper proposes a deep attention image segmentation algorithm based on neural architecture search.This method is based on the neural architecture search and explores the automatic construction of the attention module.At the same time,a new synchronous search strategy is proposed,which is used to search for the most suitable structure for different attention modules in the model,and to provide targeted optimization for different features of the network.In this article,a large number of experimental analysis of the above methods are carried out on a variety of medical data sets with different imaging parameters and imaging modes,which strongly proves its good applicability in image segmentation tasks.Visual analysis gives an in-depth discussion on the principle and interpretability of the attention mechanism.Finally,the limitations of the method and future work are summarized.