Research On Bi-Directional Constrained Semi-Supervised Segmentation Algorithm Based On Fractured Pelvis CT Images

In robot-assisted fracture surgery,the indirect visualization of the fracture site complicates the surgeon’s spatial cognition of the pelvic anatomical structure,increasing the difficulty of reduction surgery.Accurate semantic segmentation results of pelvic CT can provide supplementary information on pathology and anatomy.Therefore,the automatic segmentation of preoperative pelvic CT is of great significance for fracture reduction surgery planning and robot assisted intraoperative treatment.However,the contrast between the fracture edge and tissues is extremely low,and the local texture features between different tissues are very similar.In addition,manual annotation of CT images is time-consuming and subjectively influenced,making it extremely difficult to obtain a large amount of high-quality annotation data.The purpose of this study is to improve the segmentation accuracy of pelvic CT images by using sufficient unlabeled data,and meet the clinical requirements for pelvic structure,realistic data requirements.The specific research results are as follows:(1)In order to solve the problem of data scarcity and avoid overfitting in the training of small dataset,an interpolation consistency module based on Beta distribution is studied,which uses the pixel-level interpolation perturbation to augment the diversity of data.A pseudo label assisted supervision module is designed to constrain misclassification behavior of interpolation perturbations by learning the underlying information of unlabeled data.In addition,an effective pseudo supervised loss algorithm was discussed to ensure the fitting of the training process of the pseudo supervised module.The interpolation module and pseudo supervision module pursue consistent prediction of the same unlabeled image from the perspectives of interpolation data and underlying data,thereby improving segmentation accuracy.(2)To address the issues of low contrast and blurred edges in pelvic CT,a bi-directional constrained semi-supervised segmentation model PICT for pelvic CT was proposed.The augmented data generated by the interpolation module limits the unstable development of early pseudo label quality due to data scarcity.The pseudo supervised auxiliary module constrains the invalid disturbance of the interpolation module.PICT joints the interpolation and pseudo supervision,which are bi-constrained and complementary modules.By encouraging the predictions of teacher model and student model to be consistency at pixel,model and task level to capture more detailed structural features of data,so as to solve the problem of low contrast between soft and hard tissues of pelvic CT and fuzzy fracture edges.(3)In response to the lack of an open-source pelvic CT dataset containing anatomical structures in the field of pelvic fractures,a multi-tissues pelvic dataset annotated by experts,including pelvic structures was established.The dataset contains 100 CT slices.The muscles,tissues,and bones with extremely low contrast are classified into 7 categories.(4)PICT achieves the advanced performance of three challenging medical datasets: the 2D left atrial MRI dataset ACDC,the 3D pelvic fracture CT dataset CTPelvic1 k,and the 2D pelvic CT dataset Multi-tissue.Under the experimental conditions of using 10%,25%,and 50% labeled data on these three datasets,the average DSC scores of PICT reached 87.18%,96.42%,and79.41%,respectively.Compared with the state-of-the-art semi-supervised methods,PICT has improved by approximately 0.8%,0.5%,and 1%,respectively.Compared with the supervised method,PICT has improved by3%-9%.The progressiveness in different modes(CT/MRI),different types(2D/3D),and different objects(left atrium/pelvis)shows that the proposed algorithm has generalization and robustness.