Study Of Multi-phase Enhanced CT Image Synthesis-assisted Abdominal Multi-organ Segmentation Method

In the customization of radiation treatment planning,CECT(Contrast-enhanced computer tomography)scans are used for reliable diagnostic purposes due to the increased density difference between diseased and normal tissue after contrast injection,enhancing the brightness of the diseased area,whereas NECT(Non-enhanced computer tomography)is usually used to guide radiation dose calculations.Firstly,there are errors that registration cannot be resolved in the treatment planning system between NECT and CECT that affect the accuracy of radiation therapy dose calculation,although for deep learning modelling of automatic organ-threatening segmentation,multi-phase CT provides more comprehensive image information,and these errors still affect segmentation performance;secondly,the quality of the CT imaging is vulnerable to patient’s inability to adhere to long multi-phase CT scans due to unbearable pain and resting tremor;and finally,the In the field of medical image processing,it still faces the challenge of having little annotated data and much unannotated data.To address these problems,this thesis investigates the application of multi-phase enhanced CT synthesis method and segmentation to achieve multi-phase CT synthesis-assisted abdominal multi-organ segmentation.The main research work and innovations in this thesis are as follows.(1)In order to eliminate errors and random breathing differences that cannot be solved by alignment,and at the same time reduce the time between CECT scans and NECT scans and reduce patient exposure to radiation,a multi-phase CT synthesis method based on multi-headed self-attention perception is proposed,and the Transformer module introduced based on the multi-headed self-attention mechanism improves the network’s ability to capture long-distance semantic information by introducing perceptual loss to minimise the feature differences between real and synthesised images.Experimenting on a total of 526 cases of multi-phase CT private dataset,the mean MAEs of NECT,venous phase CECT and delay phase CECT synthesized from arterial phase CECT were 19.192 ± 3.381,20.140±2.676 and 22.538± 2.874 respectively,all outperforming the other image synthesis algorithms compared.The results show that the multi-phase CT synthesis method proposed in this thesis can not only effectively eliminate the unresolvable errors in inter-phase CT alignment,but also synthesize high-quality CT images to save scanning procedures and time.(2)In order to make use of the CT image information of different phases and reduce the annotation workload of doctors,the multi-phase CT synthesis-assisted abdominal multi-organ segmentation method is proposed,using the multi-phase CT images synthesized by the trained multi-phase CT synthesis model to segment 13 abdominal organs.The average DSC in the internal validation set is 0.847,which is better than other comparative segmentation methods,and the average DSC in the external validation set is 0.823.The results show that the multi-phase CT synthesisassisted abdominal multi-organ segmentation method proposed in this thesis can utilize the information in multi-phase CT images,thus improving the performance of abdominal multi-organ segmentation.(3)In order to cope with the situation that there are few annotated data and many unannotated data in medical image processing,and to make full use of the information in the unannotated data to improve the generalization performance of the segmentation method,the semi-supervised multi-phase CT synthesis-assisted abdominal multi-organ segmentation method is proposed,and the semi-supervised learning method uses contrast learning to mine the semantic information of unannotated data to improve the segmentation performance of abdominal 13 organs.The experimental data include 34 cases of labeled data and 492 cases of unlabeled data,and the average DSC in the internal validation set is 0.856,which is better than other comparative segmentation methods,and the average DSC in the external validation set is 0.827.The results demonstrate that the proposed multi-phase CT synthesis-assisted abdominal multiorgan semi-supervised segmentation method further improves the segmentation performance.