DCNet: Diversity convolutional network for ventricle segmentation on short-axis cardiac magnetic resonance images

Abstract

To accurately and simultaneously segment myocardium, left and right ventricles at the end-diastolic (ED) and end-systolic (ES) phases from short-axis cardiac magnetic resonance (CMR) images with inherent variability in appearance, shape, and location of the region of interest (ROI), we propose a diversity convolutional network (DCNet) that aims to solve ventricle under- and over-segmentation problems. DCNet is composed of three stages: the integration of diversity features, recoding of diversity features, and decoding of integrated features. To enhance the representational capacity and enrich the feature space of a single convolution, we design a diversity convolution block during the first stage. During the second stage, we design a dual-path channel attention mechanism to simultaneously select average and maximum features. In addition, we use a soft Dice loss function to assist in the network’s training. We conducted experiments on the 2017 Automated Cardiac Diagnosis Challenge (ACDC 2017), 2019 Multi-Sequence Cardiac MR Segmentation Challenge (MS-CMRSeg 2019), and 2020 Myocardial Pathology Segmentation Challenge (MyoPS 2020) datasets. We submitted our test results on the ACDC dataset to an online test platform, the proposed DCNet achieved Dice scores of 95.80%, 91.77%, and 91.57% in the left ventricle, right ventricle, and myocardium segmentation tasks, respectively. Compared with four representative networks, the proposed DCNet achieves the best results on balanced steady state free precession (bSSFP) cine sequence and late gadolinium enhancement (LGE) CMR sequences on the MS-CMRSeg and MyoPS datasets. Therefore, the proposed method is promising for automatic ventricle segmentation in clinical applications. We uploaded the code to https://github.com/fly1995/DCNet.

Introduction

Cardiac magnetic resonance (CMR) imaging [1] can provide doctors with high-resolution, high-contrast, and high-SNR images in any orientation. It also detects and monitors cardiovascular diseases as a non-invasive technique [2]. Cardiac function and anatomy can be evaluated using an ejection fraction from the left and right ventricles, the volume at the end-diastolic (ED) phase of a ventricle, the volume at the end-systolic (ES) phase of the myocardium, and mass at the ED phase [3]. The calculation of these clinical metrics is based on the depiction of the cardiac chamber; thus, an accurate and effective ventricle segmentation method can assist clinical diagnosis.

Before deep learning became popular, medical image segmentation methods based on a level set [4] and variational model [5] are more common and, most CMR segmentation methods were semi-automatic, which are based on active contours, graph cuts, and some atlas fitting strategies [6], [7], [8], [9]. With the rise of deep learning [10], [11], end-to-end deep learning segmentation methods [12], [13], [14], [15], [16] in conjunction with traditional segmentation methods have been frequently used. The combination of level-set [17], [18] and deep learning or deep learning and Markov random fields (MRFs) [19] are two conventional problem-solving approaches. For example, Ngo et al. [20] adopted a deep belief network to initialize contours and then segmented left ventricles with a level-set model. Avendi et al. [21] used a convolutional neural network (CNN) to initialize a level-set for left ventricle segmentation. Simantiris et al. [22] segmented CMR images using a dilated CNN with MRF model optimization.

Recently, a standalone CNN has become a popular manner of segmenting cardiac structures from CMR data. For example, Poudel et al. [23] adopted a recurrent fully CNN for multi-slice CMR segmentation. Li et al. [24] designed a multiscale dual-path feature aggregation network to segment multi-sequence CMR. Li et al. [25] adopted dilated-inception network to segment right ventricles. Indeed, these solutions for CMR segmentation have made considerable contributions [26], [27], [28], [29], [30], [31], [32], [33], [34], [35], [36], [37] to the field of medical image processing. Now, the focus on truly providing simple and effective solutions to clinicians. Therefore, this study focuses on the transferability of the method, not its complexity.

In this study, to solve ventricle over- and under-segmentation and to accurately and simultaneously segment left ventricles, right ventricles, and left ventricular myocardium at the ED and ES phases from short-axis CMR, we propose a diversity convolutional network (DCNet). Previous studies have lacked an understanding of the essential features in CMR images. In DCNet, the diversity convolution block can capture various receptive fields and feature spaces of original images, which helps solve over- and under-segmentation problems effectively. The dual-path channel attention mechanism can integrate the average and maximum features of the integration stage of diversity features, which also helps solve over- and under-segmentation problems effectively. The main contributions of DCNet are summarized as follows:

(1) We proposed an easily reproducible DCNet that comprises three stages: the integration of diversity features, recoding of diversity features, and decoding of integrated features.

(2) To enhance the representational capacity and enrich the feature space of a single convolution, we designed a diversity convolution block (DCB) that can be used as a convolution layer during the integration stage.

(3) For the recoding stage, we designed a dual-path channel attention mechanism (DCAM) to simultaneously select average and maximum feature.

(4) We adopted a soft Dice loss function to assist the network’s training. It considers entire cardiac structures and sub-structures. We conducted sufficient experiments on three datasets: 2017 Automated Cardiac Diagnosis Challenge (ACDC 2017), 2019 Multi-Sequence Cardiac MR Segmentation Challenge (MS-CMRSeg 2019), and 2020 Myocardial Pathology Segmentation Challenge (MyoPS 2020).

The remainder of this paper is organized as follows. Section 2 summarizes the challenge of automatic cardiac diagnosis. Section 3 presents the proposed method. Sections 4 Experiment results, 5 Discussion present the results and discussions, respectively. Finally, Section 6 draws conclusions.