Source-linked AI summary
Automatic 3D Cardiovascular MR Segmentation with Densely-Connected Volumetric ConvNets
Lequan Yu, Jie-Zhi Cheng, Qi Dou, Xin Yang, Hao Chen, Jing Qin, Pheng-Ann Heng
TL;DR
Accurate 3D cardiac MR segmentation is clinically important but difficult because borders can be ambiguous and anatomy varies across subjects. The paper introduces DenseVoxNet, a densely connected 3D fully convolutional network with feature reuse and auxiliary side paths. It achieves the best reported Dice performance in the challenge comparison and outperforms other 3D ConvNets with fewer parameters.
Problem
Automatic cardiac and vascular segmentation from 3D cardiac MR images is important for cardiovascular diagnosis and treatment but is challenged by ambiguous borders, anatomical variation, and costly manual labeling.
Method
DenseVoxNet uses a densely connected 3D fully convolutional architecture with concatenated feature reuse and auxiliary side paths for volume-to-volume segmentation.
Results
0.821±0.041 Dice for myocardium, the challenge ranking metric, was the best performance in the comparison, while DenseVoxNet also outperformed other 3D ConvNets with fewer parameters.
Takeaways & Limitations
DenseVoxNet provides parameter-efficient 3D cardiac and vascular segmentation and achieved the best Dice scores for myocardium and blood pool on the challenge dataset.
Takeaways & Limitations
The paper notes that summation-based identity and transformed-feature skip connections may impede information flow in conventional residual connections.
Abstract
from arXiv · showhide
Automatic and accurate whole-heart and great vessel segmentation from 3D cardiac magnetic resonance (MR) images plays an important role in the computer-assisted diagnosis and treatment of cardiovascular disease. However, this task is very challenging due to ambiguous cardiac borders and large anatomical variations among different subjects. In this paper, we propose a novel densely-connected volumetric convolutional neural network, referred as DenseVoxNet, to automatically segment the cardiac and vascular structures from 3D cardiac MR images. The DenseVoxNet adopts the 3D fully convolutional architecture for effective volume-to-volume prediction. From the learning perspective, our DenseVoxNet has three compelling advantages. First, it preserves the maximum information flow between layers by a densely-connected mechanism and hence eases the network training. Second, it avoids learning redundant feature maps by encouraging feature reuse and hence requires fewer parameters to achieve high performance, which is essential for medical applications with limited training data. Third, we add auxiliary side paths to strengthen the gradient propagation and stabilize the learning process. We demonstrate the effectiveness of DenseVoxNet by comparing it with the state-of-the-art approaches from HVSMR 2016 challenge in conjunction with MICCAI, and our network achieves the best dice coefficient. We also show that our network can achieve better performance than other 3D ConvNets but with fewer parameters.
1 Introduction
3D cardiac MR segmentation is clinically valuable but difficult because manual labeling is laborious and variable, while cardiac borders and anatomy vary substantially across subjects. The paper proposes DenseVoxNet to address these challenges with a densely connected volumetric ConvNet designed for limited training data.
- Clinical motivation: 3D cardiac-structure segmentation supports cardiovascular diagnosis, treatment planning, and patient-specific heart models for surgery.Manual segmentation of every MR slice is tedious, time-consuming, and subject to inter- and intra-observer variability.
- Segmentation challenge: Ambiguous cardiac borders caused by low contrast and substantial inter-subject anatomical variation make automatic segmentation challenging.These challenges also affect atlas-based and deformable-model approaches, which require suitable atlases from relatively large datasets.
- Proposed approach: DenseVoxNet is a densely connected volumetric ConvNet for cardiac and vascular segmentation in 3D cardiac MR images.Its 3D fully convolutional architecture supports volume-to-volume prediction and is intended to ease training with limited data.
2 Method
DenseVoxNet combines dense connectivity, 3D fully convolutional processing, down- and up-sampling, and auxiliary paths for efficient volumetric segmentation. Its design reuses features, improves information and gradient flow, and limits parameter growth for training with constrained medical datasets.
- 2.1 Dense Connection: Dense connectivity links each layer to all subsequent layers, giving layers direct supervision and encouraging feature reuse.The connections concatenate preceding feature maps rather than combining them by summation.
- 2.1 Dense Connection: The growth rate k can be set small because feature reuse reduces the need to relearn redundant feature maps, lowering parameter requirements for limited training data.The reported growth rate is k = 12.
- 2.2 The Architecture of DenseVoxNet: An auxiliary long skip connection links the transition layer to the output through a deconvolution layer, strengthening gradient propagation and stabilizing learning.The method also uses dropout following each convolutional layer to improve robustness against overfitting.
- Training procedure: Training uses randomly cropped 64×64×64 sub-volumes to fit the limited 12GB GPU memory, with final predictions obtained by major voting.The network is implemented in Caffe and optimized with stochastic gradient descent.
3 Experiments and Results
DenseVoxNet was evaluated on the limited, low-quality HVSMR 2016 dataset and achieved strong qualitative and quantitative segmentation results for myocardium and blood pool. It outperformed competing approaches, including other 3D ConvNets, while using fewer parameters and benefiting from auxiliary side paths.
- Dataset: The HVSMR 2016 evaluation used 10 training and 10 testing 3D cardiac MR scans from patients with congenital heart diseases.The scans were acquired with a 1.5T scanner and included annotations for myocardium and great vessel.
- Qualitative Results: DenseVoxNet successfully demarcated myocardium and blood pool across subjects despite large anatomical variation and low-intensity contrast.Figure 2 shows representative training and testing results, with segmentation results distinguished from available ground truth by color.
- Comparison with Other Methods: 0.821±0.041 Dice for myocardium and 0.931±0.011 Dice for blood pool were the best scores achieved by DenseVoxNet on the challenge dataset.The myocardium result outperformed the second-ranked method by around 2%, while ancillary boundary measures were also competitive.
- Comparison with Other Methods: Myocardium segmentation was more challenging than blood-pool segmentation because of ambiguous borders in low-resolution MR images.DenseVoxNet was especially better than two other ConvNet-based approaches on myocardium Dice, despite close blood-pool scores.
- Comparison with Other Methods: DenseVoxNet outperformed 3D U-Net and VoxResNet with fewer parameters, while auxiliary side paths further improved segmentation, especially for myocardium.The comparison attributes the advantage to improved information flow from dense connections and feature-map reuse under limited training data.
4 Discussion and Conclusion
DenseVoxNet addresses cardiac-structure segmentation in 3D MR images using dense connectivity and feature reuse. The reported results show strong Dice performance on the challenge dataset and advantages over other 3D ConvNets with fewer parameters.
- Discussion and Conclusion: DenseVoxNet uses dense connectivity to reuse learned features, improving parameter efficiency and reducing the chance of overfitting with limited training data.The network also uses many convolutional layers in its downsampling path to obtain large receptive fields and higher-level features.
- Discussion and Conclusion: DenseVoxNet achieved the best Dice scores for myocardium and blood pool on the challenge dataset and outperformed three 3D ConvNets.The paper notes that limited data makes effective 3D-network training difficult, while DenseVoxNet maintains better segmentation performance.