[Paper Review] Automatic 3D Cardiovascular MR Segmentation with Densely-Connected Volumetric ConvNets

Lequan Yu, Jie‐Zhi Cheng+5 more|arXiv (Cornell University)|Aug 2, 2017

Advanced Neural Network Applications18 references50 citations

TL;DR

Introduces DenseVoxNet, a densely-connected 3D fully convolutional network for automatic segmentation of myocardium and blood pool in 3D cardiac MR images, achieving state-of-the-art Dice scores with fewer parameters.

ABSTRACT

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.

Motivation & Objective

• Address the challenge of automatic whole-heart and great vessel segmentation in 3D cardiac MR images with limited training data.
• Propose a densely-connected volumetric CNN to improve information flow and reduce parameters.
• Enhance training stability and gradient propagation with auxiliary side paths and deep supervision.
• Demonstrate superiority over state-of-the-art 3D ConvNets on HVSMR 2016 dataset in myocardium and blood pool segmentation.

Proposed method

• Proposes DenseVoxNet, a 3D fully convolutional network with dense connectivity to enable feature reuse and reduced parameter count.
• Implements two DenseBlocks with 12 transformation layers each, using a growth rate of 12.
• Uses 3×3×3 convolutions, batch normalization, ReLU, and dropout (rate 0.2) along the downsampling path.
• Incorporates a transition block between DenseBlocks and a two-stage upsampling path with deconvolutions to produce voxel-wise segmentations.
• Adds a long skip connection from the transition layer to the final output to foster deep supervision and gradient flow.
• Trains on sub-volumes 64×64×64 due to memory constraints, with major voting over overlapped sub-volume predictions for final segmentation.

Experimental results

Research questions

• RQ1 Can a densely-connected 3D CNN improve automatic segmentation of myocardium and blood pool in 3D cardiac MR images with limited training data?
• RQ2 Does DenseVoxNet achieve better segmentation performance with fewer parameters than existing 3D ConvNets (e.g., 3D U-Net, VoxResNet)?
• RQ3 Do auxiliary side paths and deep supervision further enhance segmentation accuracy and training stability?

Key findings

• DenseVoxNet achieves the best Dice scores for both myocardium (0.821 ± 0.041) and blood pool (0.931 ± 0.011) on the HVSMR 2016 dataset among competing methods.
• DenseVoxNet uses about 1.8M parameters, substantially fewer than 3D U-Net (19.0M) and VoxResNet (4.0M).
• DenseVoxNet without auxiliary paths (DenseVoxNet-A) underperforms compared to the full DenseVoxNet, illustrating the benefit of auxiliary deep supervision.
• Compared to other ConvNet approaches, DenseVoxNet achieves higher myocardium Dice while maintaining competitive or better blood pool Dice, with fewer parameters.
• The network demonstrates improved information flow and reduced overfitting risk due to dense connectivity and feature reuse.