Research On Multidimensional Placenta Information Processing And Quantitative Analysis

As the bridge between the fetus and the mother,the placenta is vital in the growth of the fetus.The placenta itself not only reflects the health status of the mother and baby,but its gross and microscopic characteristics are also associated with diseases.In recent years,computer vision algorithms have been developed and have achieved fruitful results in the medical field,bringing new opportunities for the study of the placenta,especially the pathology of the postpartum placenta.The gross examination is the first stage of the postpartum placenta,recording basic information such as shape,colour and lesions.The current examination is time-consuming and labour-intensive,with only written reports remaining.Due to differences in individual perception and the characteristics of human recording,the content of the report is inevitably subjective and tends to be more descriptive than quantitative.In addition,text is far less intuitive than images and digital models,with much placental information being lost.All of these issues are detrimental to the preservation of placental data and subsequent research.This thesis focuses on the quantitative analysis of placenta gross examination.A multi-dimensional information collection and analysis system is designed to process the acquired 2D images and 3D point clouds with segmentation,localization and alignment algorithms,obtaining geometric,localization and 3D quantitative medical indicators which assist physicians to perform gross examination quickly and efficiently.The research in this thesis is divided into three main aspects:1.To tackle the issue that distillation algorithms have difficulty transferring effective information in the middle layers of networks with different architectures,the thesis proposes a multi-scale Multi-KD distillation algorithm based on semantic segmentation.The algorithm takes Res Block-UNet,with good segmentation accuracy,as the teacher network and Atta Net,with less computational resources and faster inference,as the student network,effectively improving the segmentation accuracy by supervising the output of Atta Net at multiple scales,and improving the IOU of umbilical cord segmentation by 5%,making it meet the dual requirements of accuracy and speed in practical applications.2.To tackle the issue of fuzzy edge and the similarity of texture and background,this thesis proposes a localization algorithm based on skeleton extraction.The algorithm makes full use of the prior information that the attachment point must be the endpoint of the umbilical cord.Through detection of the endpoint,the problem of identifying the attachment point in the full image is transformed into a classification problem of determining whether the endpoint is the attachment point,and then the final result is obtained with the classifier.The algorithm can achieve an average accuracy of 77.81%.This thesis also achieves better localization performance by combining the respective strengths of this algorithm and Yolo v5 to improve the accuracy to 85.07%.3.To tackle the issue that the placenta point cloud is difficult to acquire key points and the alignment is time-consuming,this thesis proposes a SAC-IA alignment method based on convex polygon boundary points,which selects a few high-quality key points for alignment.The algorithm effectively exploits the properties of the placenta point cloud data to reduce the time required for coarse alignment with a mean square error of1.08,which means it is also advantageous in terms of alignment accuracy.Based on that,a complete point cloud of the upper surface of the placenta is obtained by fine alignment KD-ICP,and a digital model of the placenta is obtained by surface reconstruction with Poisson algorithm.In addition,this thesis analyzes the geometric,localization and 3D medical indicators of the placenta based on the results of segmentation,localization and alignment algorithms.In conclusion,the placenta gross collection and analysis system designed in this thesis can not only acquire and store multi-dimensional placenta information,but also process 2D and 3D data through segmentation,localization and alignment algorithms to achieve quantitative analysis of medical indicators,which is of great significance to improve the efficiency of placenta gross examination and objectivity of examination results.