Research On Medical Image Segmentation Algorithm

At present,computer technology is developing rapidly.With the updating of medical equipment,medical image data is increasing.The processing and analysis of medical images can assist,enlighten and promote the diagnosis and treatment of doctors.Medical image segmentation is the basic technology of medical image processing.It can mark regions of interest in medical images at the pixel level based on application requirements.The imaging methods of medical images are complicated,and their characteristics are also different.At the same time,medical images obtained by different imaging methods may be targeted at different parts of the human body.Therefore,medical image segmentation tasks need to build different models for data from different imaging methods.This paper focuses on the study of three medical image segmentation models.These three types of images include: wound images,pathological images,and high-resolution computer tomography(HRCT)images.Their target organs are different and their characteristics vary widely.The main contributions of this paper are as follows.1.The main problems of wound segmentation include:(1)The background of the image is complex and will lead to incorrect segmentation results.(2)Difficulties in image acquisition of wounds lead to less data,which is not conducive to the training of deep neural networks.Aiming at these problems,this paper proposes a wound segmentation method for complex backgrounds.Based on the semantic information of the wound,the model uses traditional methods to improve the segmentation accuracy of deep neural networks.The basic idea is:(1)Based on the morphological methods and the skin features,an algorithm for detecting skin with wound is designed,based on which the complex background is excluded.(2)A multi-background training method is designed.The model is trained with a set of wound images containing and removing complex backgrounds.During the test,the model segmented the wound image without the complex background.This method keeps the generalization performance of the model while reducing the interference factors in the process of wound identification.(3)Based on the wound semantics,a refinement algorithm for wound semantics was designed.The algorithm improves the effect of wound segmentation by eliminating unreasonable regions in the segmentation results.Experimental results verify the effectiveness of the model.2.The deep neural networks for image segmentation have the following problems:(1)The pooling and convolution operations of the deep neural networks will blur the location information of images and affect the accuracy of the image segmentation.(2)Deep neural networks cannot use the image semantic information directly,which will cause errors in the segmentation results.For these difficulties,this paper proposes a wound segmentation network based on location information enhancement,which effectively improves image segmentation accuracy by enhancing image location information and semantic correction based on the smooth convolution kernel.The basic idea is:(1)An image location enhancement method is proposed.The two-dimensional location information of the image is represented as an location map.The location map is added to the model training,which can effectively improve the network’s ability to extract image location information.(2)A location enhanced convolution method is designed.The model uses convolution kernels with empirical values to highlight the positional relationship between image pixels and improve the efficiency of the use of position information.(3)A segmentation smoothing method based on semantic correction is designed.The smooth convolution kernel is designed according to the image semantic features,and the segmentation result is corrected to reduce the errors.The experimental results show that the proposed method can improve the network’s ability to extract the location information of the wound and optimize the network segmentation results.3.The problems in the segmentation of gastric adenocarcinoma pathology include:(1)There is an area of intersection between cancer tissue and healthy tissue,resulting in blurred edges and difficulty in segmentation.(2)The size of the pathological image is too large,which makes it impossible to use deep neural networks to segment the pathological image directly.Aiming at these problems,this paper proposes a cascaded and refined pathological image segmentation method.The model transforms the image segmentation into a step-by-step refinement block classification problem.The model obtains segmented coarse labels by reducing the size of the blocks used for classification,and trains a segmentation network based on these labels.The basic idea is:(1)A pathological block classification network based on medical observation is proposed.According to medical observations,the model cuts the whole slide image into pathological patches of empirical size for class labeling.These labels are used to train a stable block classification network.This method solves the difficulty of labeling pathological image segmentation.(2)A method of subdivision of pathological block based on tissue environment was proposed.The model cuts the pathological images with overlap to obtain pathological blocks.The classification results of overlapping regions are determined based on the classification information of each block,thereby completing the classification of smaller pathological blocks.(3)Based on the above method,a cascading and refined pathological image segmentation framework is designed.Based on the classification result of the fine block,the image is segmented and labeled,and the pixel-level image segmentation model is obtained.Experiments show that the model can solve the problem in the task well and can complete the precise segmentation task of the whole slide image.4.The chest HRCT image of interstitial lung disease has the following problems in the lung parenchymal segmentation task:(1)The environmental noise in the image will affect the segmentation results of the lung parenchyma.(2)There are abnormal areas in the lungs due to illness.These abnormalities are similar to those of non-lung tissues and are difficult to segment.Aiming at these problems,this paper proposes an image segmentation model based on watershed and morphological methods.The basic idea is:(1)Based on the semantic features of the thoracic region,a thoracic segmentation model was constructed using morphological methods.The chest region is segmented to eliminate the interference of the environmental background.(2)A method of segmentation of lung parenchyma based on multiple segmentation correction is proposed.The watershed method is used to segment the lung parenchymas with multiple resolutions.Then,the difference of the segmentation results are abtained.Based on the classification of the difference,the markers of abnormal regions in the watershed are corrected to obtain the lung parenchyma.The precise segmentation solves the problem of mis-segmentation of abnormal regions in the lung parenchyma.The experimental results show that the model can well solve the influence of background and special abnormal regions on the segmentation of lung parenchyma.