Automatic localization of anatomical landmarks in cardiac MR perfusion using random forests

Highlights

• A method for automatic landmark localization in cardiac DCE MR perfusion is proposed.

• Automatic segmentation of the RV cavity is robust and helps to reduce the search space.

• Random forest classifiers are used to distinguish landmarks and non-landmarks for LV center and RV insertion point.

• Mean localization errors of the proposed method were lower than those of the histogram of oriented gradient method.

Abstract

Automatic localization of anatomical landmarks in myocardial perfusion magnetic resonance (MR) data is considered to be a preliminary step toward fully automatic quantification of regional upslope or blood flow in the myocardium. The goal of this work is to develop an automatic method based on supervised learning and detect anatomical landmarks with high accuracy from myocardial perfusion MR data. Dynamic contrast enhanced myocardial MR perfusion data were acquired using a standard perfusion sequence. To effectively extract characteristic features for left ventricle (LV) center point and anterior right ventricle (RV) insertion point, we performed feature extraction of the two landmarks independently from an LV enhancement frame and an RV enhancement frame. Feature extraction of pixel intensity, Sobel gradient, and Haar-like features was performed. Cross-validation from training data was used to build a random forests classifier. We used 38 subjects’ data as training datasets and 21 subjects’ data as test datasets. The proposed method provided high accuracy in localization of the LV center point and anterior RV insertion point. The mean (±SD) localization errors of the proposed method were 3.38 (±2.36) mm for the LV center point and 4.23 (±1.97) mm for the anterior RV insertion point. The proposed method shows the potential to automatically localize anatomical landmarks for the segmental analysis of myocardial perfusion in MRI.

Introduction

Automatic localization of anatomical landmarks is an important preliminary step toward fully automatic processing of myocardial perfusion MR data. In myocardial MR perfusion analysis [1], anatomical landmarks can be the LV center point and RV anterior insertion point (see Fig. 1a). After the identification of these locations, one can divide the myocardium into regional myocardial segments [2] (see Fig. 1a) and obtain myocardial perfusion indices (e.g., myocardial blood flow, upslope) in each myocardial segment (see Fig. 1b) [3], [4]. This segmental analysis helps regional quantitative assessment of myocardial ischemia which is likely to result from stenosis of one of three main coronary arteries [1].

Landmark localization in cardiac MR perfusion has not been a primary topic of research, and the procedures of localizing the RV insertion point were introduced in the literature. Weng et al. found the RV insertion point by searching for the uppermost pixel in the right ventricular blood pool that touches the myocardium [5]. However, this method relied on the performance of myocardial segmentation in which automatic and accurate segmentation of the myocardium in cardiac MR perfusion is a challenging task [6]. In Tarroni et al. [7], the user manually identified the anterior point of insertion of the RV free wall into the interventricular septum. This manual procedure is tedious and is subject to intra or inter-subject variabilities. Automatic selection of the landmarks would be beneficial since it is objective and serves as a sub-procedure for fully automating the regional quantification of myocardial perfusion.

Several research groups have investigated supervised learning based approaches for automatic detection and localization of anatomical landmarks in a variety of biomedical imaging applications [8], [9], [10], [11], [12], [13]. Vandaele et al. extracted a vector of visual features at different resolutions for each labeled pixel and used a randomized forests algorithm to accurately detect the landmarks in unseen X-ray cephalometry images [9]. Stern et al. used a random forests classifier [14] and applied it to zebrafish images [10]. Lu et al. adopted a joint context based approach to automatically detect the apex and the two basal annulus points from a single cardiac cine MR long axis slice [11]. Sedai et al. extracted multiscale histogram of oriented gradients (HOG) descriptors and applied a random forests classifier to RV landmark detection in cardiac cine MR data [12]. Mahapatra used a combination of graph cut segmentation, low level features, and random forests classifier to detect landmarks in long and short axis orientations of cardiac cine MR data [13].

In this work we apply a supervised learning approach based on random forests to myocardial MR perfusion data. To our knowledge, automatic localization of anatomical landmarks using machine learning has not been demonstrated in cardiac perfusion MR dataset in the literature. As myocardial MR perfusion data contain multiple frames reflecting the sequential dynamic contrast enhancement in the RV blood pool, and LV blood pool, followed by the myocardium, we obtain features for RV insertion point detection from an RV enhancement frame and features for LV center point detection from an LV enhancement frame, respectively. We present image processing procedures for feature extraction and assess cross-validation results of the random forests model. Finally, we demonstrate the performance of model prediction in detecting the two landmarks in 21 test subjects.