DCNet: Diversity convolutional network for ventricle segmentation on short-axis cardiac magnetic resonance images
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.
Section snippets
Previous work
The cardiac segmentation challenge has been held annually [50], [51], [52], [53]. To the best of our knowledge, ACDC is the only challenge that depicts the left ventricular myocardium, left ventricle, and right ventricle at the ED and ES phases in 150 patients. The active website and online test platform for the competition remain open. The ACDC was held in conjunction with the 20th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI). The data of this
Proposed network
Fig. 1 illustrates the proposed DCNet for ventricle segmentation on short-axis CMR images. The network is used to segment the left ventricle, left ventricular myocardium, and right ventricle in the ED and ES phases. DCNet comprises three stages: the integration of diversity features, recoding of diversity features, and decoding of integrated features. During these three stages, the DCB can be equivalent to a single convolution layer for deployment. DCB aims to capture diversity features of the
Dataset and data processing
We conducted experiments on the ACDC 2017 [64], MS- CMRSeg 2019 [53], [65], and MyoPS 2020 [53], [65] datasets. We adopted 80 ACDC training datasets for training and 20 ACDC training datasets for testing in the ablation experiments. In addition, we used all 100 and 50 patients for training and testing, respectively. We adopted 45 MS-CMRSeg cases for training and 25 MyoPS cases for testing to verify the generalization ability of the proposed DCNet. We integrated these five classifications into
Discussion
In Fig. 1, we use to represent the dilation rate of the dilation convolution, where is an integer greater than 0 and determined by ablation experiments. The ablation experiments were conducted with 80 patients from the ACDC training data for training and 20 patients from the ACDC training data for testing. In Table 7, Proposed+ k=[1-5] achieves the highest Dice score in all ventricle segmentation tasks. Therefore, we used Proposed+ k=[1-5] as the final model in the segmentation tasks of the
Conclusion
In this study, we propose DCNet for ventricle segmentation. To enhance the representational capacity and enrich the feature space of a single convolution, the proposed DCNet comprises DCBs. In the input stage of DCNet, we integrate multiple DCBs to obtain more features of the target area. We adopted a soft Dice loss function that quantifies the myocardium, left ventricle, right ventricle, and entire ventricle to assist in network’s training. We conducted experiments using the ACDC 2017,
CRediT authorship contribution statement
Feiyan Li: Methodology, Validation, Visualization, Writing – original draft, Writing – review & editing. Weisheng Li: Funding acquisition, Project administration, Resources. Xinbo Gao: Resources. Rui Liu: Investigation, Supervision. Bin Xiao: Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the National Key Research and Development Program of China [2019YFE0110800], National Natural Science Foundation of China [61972060, 62027827], Natural Science Foundation of Chongqing [cstc2020jcyj-zdxmX0025, cstc2019cxcyljrc-td0270, cstc2019jcyj-cxttX0002], and Innovative Talents Program for Doctoral students of Chongqing University of Posts and Telecommunications, China [BYJS202110].
References (68)
- et al.
Standardized cardiovascular magnetic resonance imaging (CMR) protocols, society for cardiovascular magnetic resonance: board of trustees task force on standardized protocols
J. Cardiovascul. Magn. Reson.
(2008) - et al.
Graph cut segmentation with a statistical shape model in cardiac MRI
Comput. Vis. Image Underst.
(2013) - et al.
A review of segmentation methods in short axis cardiac mr images
Med. Image Anal.
(2011) - et al.
Unified medical image segmentation by learning from uncertainty in an end-to-end manner
Knowl.-Based Syst.
(2022) - et al.
A shape-guided deep residual network for automated CT lung segmentation
Knowl.-Based Syst.
(2022) - et al.
MallesNet: A multi-object assistance based network for brachial plexus segmentation in ultrasound images
Med. Image Anal.
(2022) - et al.
Deep U-net architecture with curriculum learning for myocardial pathology segmentation in multi-sequence cardiac magnetic resonance images
Knowl.-Based Syst.
(2022) - et al.
Combining deep learning and level set for the automated segmentation of the left ventricle of the heart from cardiac cine magnetic resonance
Med. Image Anal.
(2017) - et al.
A combined deep-learning and deformable-model approach to fully automatic segmentation of the left ventricle in cardiac MRI
Med. Image Anal.
(2016) - et al.
MDFA-Net: Multiscale dual-path feature aggregation network for cardiac segmentation on multi-sequence cardiac MR
Knowl.-Based Syst.
(2021)
Information aggregation and fusion in deep neural networks for object interaction exploration for semantic segmentation
Knowl.-Based Syst.
Pixel-wise triplet learning for enhancing boundary discrimination in medical image segmentation
Knowl.-Based Syst.
MOLS-Net: Multi-organ and lesion segmentation network based on sequence feature pyramid and attention mechanism for aortic dissection diagnosis
Knowl.-Based Syst.
MH-Net: Model-data-driven hybrid-fusion network for medical image segmentation
Knowl.-Based Syst.
Improving cardiac mri convolutional neural network segmentation on small training datasets and dataset shift: A continuous kernel cut approach
Med. Image Anal.
AdaEn-Net: An ensemble of adaptive 2D–3D fully convolutional networks for medical image segmentation
Neural Netw.
Right ventricle segmentation from cardiac MRI: a collation study
Med. Image Anal.
An automatic cardiac segmentation framework based on multi-sequence MR image
The gold standard for noninvasive imaging in coronary heart disease: magnetic resonance imaging
Curr. Opin. Cardiol.
A review of heart chamber segmentation for structural and functional analysis using cardiac magnetic resonance imaging
Magn. Reson. Mater. Phys. Biol. Med.
A level set approach to image segmentation with intensity inhomogeneity
IEEE Trans. Cybern.
A variational approach to simultaneous image segmentation and bias correction
IEEE Trans. Cybern.
Active contour-based method for finger-vein image segmentation
IEEE Trans. Instrum. Meas.
Max-flow segmentation of the left ventricle by recovering subject-specific distributions via a bound of the bhattacharyya measure
Med. Image Anal.
Deep-learning-based lossless image coding
IEEE Trans. Circuits Syst. Video Technol.
Deep learning-based technology in responses to the joint call for proposals on video compression with capability beyond HEVC
IEEE Trans. Circuits Syst. Video Technol.
JAS-GAN: generative adversarial network based joint atrium and scar segmentation on unbalanced atrial targets
IEEE J. Biomed. Health Informat.
Simultaneous extraction of endocardial and epicardial contours of the left ventricle by distance regularized level sets
Med. Phys.
Segmentation of prostate from ultrasound images using level sets on active band and intensity variation across edges
Med. Phys.
Adaptive volumetric texture segmentation based on gaussian markov random fields features
Pattern Recognit. Lett.
Cardiac MRI segmentation with a dilated CNN incorporating domain-specific constraints
IEEE J. Sel. Top. Signal Process.
Recurrent fully convolutional neural networks for multi-slice mri cardiac segmentation
Dilated-inception net: multi-scale feature aggregation for cardiac right ventricle segmentation
IEEE Trans. Biomed. Med. Eng.
Discrete iterative partial segmentation technique (DIPS) for tomographic reconstruction
IEEE Trans. Comput. Imag.
Cited by (7)
FRBNet: Feedback refinement boundary network for semantic segmentation in breast ultrasound images
2023, Biomedical Signal Processing and ControlCAT-Seg: cascaded medical assistive tool integrating residual attention mechanisms and Squeeze-Net for 3D MRI biventricular segmentation
2024, Physical and Engineering Sciences in MedicineMulti-OCDTNet: A Novel Multi-Scale Object Context Dilated Transformer Network for Retinal Blood Vessel Segmentation
2023, International Journal of Pattern Recognition and Artificial IntelligenceSegmentation of the Left Ventricle by SDD double threshold selection and CHT
2023, Proceedings - International Conference on Image Processing, ICIP