Research On Medical Ultrasound Image Segmentation Algorithm Based On Simple Interaction

With the increasing demand for medical treatment,it is of great significance to quickly achieve accurate segmentation of medical ultrasound images to improve the quality and efficiency of clinical diagnosis,and it is of great application value and social benefit to study medical ultrasound image segmentation algorithms to assist medical personnel to intelligently segment and analyze medical ultrasound images.In this thesis,after summarizing the existing medical ultrasound image segmentation methods and investigating the specific requirements of clinical scenarios in hospital imaging departments,we propose a medical ultrasound visceral tumor image segmentation network based on multi-scale feature extraction,and propose a medical ultrasound image segmentation algorithm based on heat map interaction based on this network,and design and implement an interactive medical ultrasound image segmentation and analysis system based on the above algorithms to meet the needs of clinicians to the maximum extent.The works in this thesis mainly include:(1)A medical ultrasound image segmentation network based on multi-scale feature extraction is proposedThe network is based on a baseline approach with an encoder-decoder structure,and the baseline network is improved for more complex segmentation scenarios to obtain better segmentation performance.Specifically,firstly,the feature extraction capability of the network is enhanced by introducing an effective channel attention module for feature recoding of high-dimensional features in the deep layer of the network;secondly,for the problem of large variability in target size and morphology in medical ultrasound visceral tumor images,the method follows and improves the pooling pyramid module in the baseline network to improve the segmentation performance of the network;finally,in the Finally,in the decoder part of this network,the feature reuse module is used to extract the high-dimensional features from the deeper layers of the network and then guide the generation of the final prediction results,which brings an improvement to the segmentation accuracy of the network.In addition,the method has deeper network layers compared to the baseline network and incorporates the corresponding jump connections.Experiments on ultrasound visceral tumor image data after intercepting the region of interest show that the network has higher segmentation accuracy compared with other ultrasound image segmentation networks on metrics commonly used in medical image segmentation.(2)A medical ultrasound image segmentation algorithm based on heat map interaction is proposedThe algorithm consists of a coarse segmentation phase and a fine segmentation phase,which are implemented by a modified Hourglass network and a medical ultrasound image segmentation network based on multiscale feature extraction,respectively.The algorithm uses three Gaussian heat maps with different standard deviations to simulate mouse clicks,together with the reduced resolution ultrasound image for coarse segmentation by the modified Hourglass network,and obtains the size information of the target and its position information in the image according to the coarse segmentation result,and uses this information to intercept the region of interest for the original resolution image,and finally refines it by the fine segmentation network in the second stage The segmentation is finally refined by the second stage of fine segmentation network and the segmentation results are restored to the original ultrasound image using the size and location information.The algorithm achieves faster segmentation speed and higher segmentation accuracy than direct segmentation of high-resolution ultrasound images by two stages under the guidance of interactive information.(3)Design and development of an interactive medical ultrasound image segmentation and analysis systemIn order to apply the above algorithm,this thesis designs and implements an interactive medical ultrasound image segmentation and analysis system according to the clinical requirements.The system integrates the algorithm of this thesis and can segment medical ultrasound images interactively.In addition,the system can automatically quantify and analyze the segmentation results,which reduces the steps of manual operation by doctors and thus improves the efficiency of clinical diagnosis.