Research On Skeletal Muscle Segmentation Algorithm Based On Low-dose CT

As an important part of human muscles,skeletal muscle plays an important role in providing body power,assisting the normal operation of organs,and protecting bones.In addition,sarcopenia due to skeletal muscle loss is considered to be an important hallmark of various diseases such as rheumatoid arthritis,recurrent falls and fractures,impaired cardiopulmonary function,and end-stage liver disease.Therefore,quantitative research on skeletal muscle has important reference value in clinical applications such as disease diagnosis and treatment plan design,disease risk control,and survival cycle prediction.Due to its non-invasive scanning method,low-dose CT has the characteristics of less radiation and easy access in daily diagnosis,and is increasingly used in the quantitative analysis of human skeletal muscle.However,the accurate quantitative analysis of skeletal muscle is still in the exploratory stage.Most of the existing quantitative analysis of skeletal muscle is based on the muscle content in a single-frame cross-sectional CT image to achieve the estimation and quantification of whole body muscle tissue.The difference is very large,and it is difficult for a fixed single-frame slice to accurately reflect the muscle distribution structure of the human body,resulting in poor prediction accuracy.Therefore,how to directly and accurately quantify human trunk skeletal muscle tissue based on the entire low-dose CT image is an urgent problem to be solved in the field of medical image analysis.In order to overcome the above difficulties,this paper firstly establishes a skeletal muscle definition system conforming to human anatomy based on the mapping characteristics of human anatomical structures in low-dose CT images,and based on this definition,a human torso composed of 50 low-dose CT images is created.Skeletal muscle dataset.Secondly,this paper designs a skeletal muscle contour segmentation model based on attention mechanism,and uses the features extracted by channel and voxel self-attention module to fuse with each other to achieve accurate segmentation of contour regions.Thirdly,this paper designs a dual-decoder skeletal muscle segmentation algorithm based on regional supervision,which uses the contour regions obtained in the previous stage to prompt the segmentation network to focus on the information extraction of skeletal muscle regions,so as to achieve accurate segmentation of skeletal muscle tissue.In addition,this paper also proposes an error sampling strategy based on the dynamic error map,which further improves the segmentation accuracy of the above two models.Finally,this paper conducts experiments on the proposed algorithm in the self-built data set,and compares it with the classic network model to verify the effectiveness of the proposed algorithm in skeletal muscle segmentation accuracy.