S-Space College of Engineering/Engineering Practice School (공과대학/대학원) Program in Bioengineering (협동과정-바이오엔지니어링전공) Theses (Master's Degree_협동과정-바이오엔지니어링전공)
REAL-TIME ESTIMATION OF THE LEFT VENTRICULAR VOLUME FROM ECHOCARDIOGRAM DURING CARDIOPULMONARY RESUSCITATION USING CONVOLUTIONAL NEURAL NETWORK
합성곱 신경망을 이용한 심폐소생술 중 심초음파 영상에서 실시간 좌심실 부피 추정
- 공과대학 협동과정 바이오엔지니어링전공
- Issue Date
- 서울대학교 대학원
- Cardiopulmonary resuscitation; Echocardiography; Segmentation; Convolutional neural network; Gated recurrent unit; Left ventricular model
- 학위논문 (석사)-- 서울대학교 대학원 공과대학 협동과정 바이오엔지니어링전공, 2017. 8. 이정찬.
- This thesis describes the method for real-time segmentation based on echocardiography and three-dimensional transformation model for the left ventricular volume estimation during cardiopulmonary resuscitation (CPR). Because all people have a different structure of thoracic and the position of the heart, it has been required to optimize CPR by a person. As one of the improved methods, bio-signal feedback using echocardiography CPR is carried out. Echocardiography shows how the heart is compressed by chest compression, which directly shows cardiac output. There are two steps in estimating the cardiac output in echocardiography. The left ventricular segmentation from the echocardiography is needed to be segmented. After that, the three-dimensional volume is required to be estimated with two-dimensional segmented images. However, echocardiography during CPR is difficult due to the instability of contact between the transducer and the chest. Moreover, the previous models that map the segmented two-dimensional image to the left ventricular volume assume the heart is contracted isometrically, which is different from the condition of the heart during CPR. To solve these problems, the method for segmentation of the left ventricle stable during CPR and the model that can be applied to CPR conditions is suggested in this dissertation. The convolutional neural network is adopted to the left ventricular segmentation problem. Based on the structure of SegNet that is a fully convolutional network for real-time segmentation, skip connection and dice coefficient are applied to adapt the model to echocardiography domain. The former one helps the network to preserve the information of original images, and the latter one is used for stable segmentation. Moreover, Gated recurrent unit that is used for time series data analysis is applied to reflect the previous frames. The network achieves robust and accurate segmentation by referencing the previous frames in the segmentation of current frame. Comparing to Geodesic Active Contour method that shows the best performance in echocardiography, the proposed algorithm accomplishes higher accuracy and robust to unclear images. The left ventricular model is derived with applying constraints during CPR for modeling problem. The heart during CPR is not contracted. Thus, the assumption of the same surface between the diastolic heart and compressed heart is used. Moreover, the single ellipsoid model with the same length in the minor and intermediate axes is adopted. In comparison experiment to ETCO2 that affects the cardiac output during CPR, the proposed model show much greater correlation than the previous model.