Direct Estimation of Cardiac Bi-ventricular Volumes with Regression Forests

Abstract

Abstract. Accurate estimation of ventricular volumes plays an essential role in clinical diagnosis of cardiac diseases. Existing methods either rely on segmentation or are restricted to direct estimation of the left ventricle. In this paper, we propose a novel method for direct and joint volume estimation of bi-ventricles, i.e., the left and right ventricles, without segmentation and user inputs. Based on the cardiac image representation by multiple and complementary features, we adopt regression forests to jointly estimate the two volumes. Our method is validated on a dataset of 56 subjects with a total of 3360 MR images which shows that our method can achieve a high correlation coefficient of around 0.9 with manual segmentation obtained by human experts. With our proposed method, the most daily-used estimation of cardiac function, e.g., ejection fraction, can be conducted in a much more efficient, accurate and convenient way.

Multi-scale deep networks and regression forests for direct bi-ventricular volume estimation

“…Direct estimation of cardiac volumes has started to generate increasing interest due to the avoidance of intermediate segmentation (Afshin et al, 2012a;Wang et al, 2014;Zhen et al, 2014d). To directly estimate the ejection fraction of the LV, global image statistics are used to calculate LV volumes in (Afshin et al, 2012a).…”

“…However, segmentation itself is an extremely challenging problem which in clinical practise physicians are not interested in. Direct estimation methods which remove the segmentation step become attractive in cardiac function diagnosis and ventricular estimation due to its efficiency and clinical significance (Afshin et al, 2012a,b;Wang et al, 2014;Afshin et al, 2014;Zettinig et al, 2014;Zhen et al, 2014d;Wang et al, 2013;Zhen et al, 2015a). A comprehensive study of methods for cardiac ventricular volume estimation has been conducted in (Zhen et al, 2014c) showing that direct estimation methods provide more accurate estimation than segmentation-based methods for both left and right ventricles (LV and RV).…”

“…In this paper, towards the full capacity of direct estimation, we propose a general framework for bi-ventricular volume estimation. A preliminary conference version of this work appeared in MICCAI 2014 (Zhen et al, 2014d). In this journal version we extend in two main aspects: 1) we replace the handcrafted features in (Zhen et al, 2014d) by feature engineering with representation learned by our multi-scale deep networks (Bengio et al, 2013).…”

“…A preliminary conference version of this work appeared in MICCAI 2014 (Zhen et al, 2014d). In this journal version we extend in two main aspects: 1) we replace the handcrafted features in (Zhen et al, 2014d) by feature engineering with representation learned by our multi-scale deep networks (Bengio et al, 2013). In contrast to feature engineering, the unsupervised learning by deep networks allows us to both make full use of plenty of unlabeled data which is largely available in medical imaging and faithfully detect data-driven features for informative image representations in specific tasks; 2) we provide a wider investiga-A C C E P T E D M A N U S C R I P T tion on a larger dataset of 100 subjects with 6000 images than previously used.…”

Unsupervised Myocardial Segmentation for Cardiac MRI

“…Segmentationonly class of supervised techniques, on the other hand, mainly focuses on featurebased representation of the myocardium. Texture information is generally considered as an effective feature representation of the myocardium for standard CINE MR images [22]. Patch-based static discriminative dictionary learning technique (SJTAD) [16] and Multi-scale Appearance Dictionary Learning technique [4] have achieved high accuracy and are considered as state-of-the-art mechanisms for supervised myocardial segmentation.…”