의료영상의 자동 폐 분할 및 분석에 관한 연구

Abstract

Due to development of medical image scanners in medical field, various modality of medical images are used in recent years. In order to diagnose and treat patients using medical image, analysis of medical image is necessary. Acquired information from analysis of medical image is used effectively in process of diagnosis and treatment of patients. In case of chest CT image, about 300~500 of CT images can be obtained from a patient. Actually, it is impossible for a radiologist to analyze all the images. Even if a radiologist analyzes all the images, there could be a serious drawback that analysis results of one patient are at risk for being interpreted differently by the workmanship of radiologists. For such a reason, automatic analysis of medical image is required. Medical image segmentation is an important stage before analyzing of medical image. Through segmentation of medical image, not only a medical specialist can easily observe changes of organ and lesion, but also receive technical help for surgery plan. There are many methods in segmentation of medical image, such as threshold, Region Growing, Watershed, ASM, Clustering, Level-set, and etc. However, these methods have problems such as long running time, required interaction of user in segmentation process, etc. Level-set is great tool for modeling, like inflation of an airbag, or a drop of oil floating in water. For such a reason, it is suitable segmentation of medical image. However, it has drawbacks that requires input of initial contour of user and takes long running time. In this thesis, we performed Level-set to segment lung regions from chest CT image and pursued a research on solutions for problems in initial contour and running time. If optimized initial contour to the shape of object is used in Level-set, repetition number of Level-set is reduced. Consequently, running time of Level-set is reduced. Using MRA, running time taken in initial contour auto setting stage was reduced. Because of data loss in MRA, CIM was suggested, and errors by data loss in initial contour setting were reduced. As a result of performing lung segmentation by proposed MRA Level, performance of existing Level-set was maintained. Testing with CT image DB in VESSEL12, more than 0.98 of average accuracy was confirmed. In order to confirm reduction of running time in initial contour setting and image segmentation, we checked necessary Cycle number using MRA. As a result, compared to 5.373×1011 of Level-set that users input initial contour by 7.390×109, we confirmed 72 times of reduction. In addition, we analyzed lung region of LIDC chest CT image and lesion region using Adaboost which is binary classifier.