[Paper Review] Shape Constrained CNN for Cardiac MR Segmentation with Simultaneous Prediction of Shape and Pose Parameters

Sofie Tilborghs, Tom Dresselaers+3 more|arXiv (Cornell University)|Oct 18, 2020

Advanced Neural Network Applications17 references8 citations

TL;DR

This paper proposes a shape-constrained CNN that jointly regresses cardiac left ventricle (LV) shape and pose parameters from a statistical shape model while simultaneously predicting segmentation via signed distance maps. By integrating shape priors through learned coefficients and enforcing consistency with segmentation via a multi-task loss, the method achieves state-of-the-art accuracy in LV and myocardial segmentation, with 99% correlation for LV area and 88% for regional wall thicknesses on an in-house dataset, and superior performance on public benchmarks.

ABSTRACT

Semantic segmentation using convolutional neural networks (CNNs) is the state-of-the-art for many medical segmentation tasks including left ventricle (LV) segmentation in cardiac MR images. However, a drawback is that these CNNs lack explicit shape constraints, occasionally resulting in unrealistic segmentations. In this paper, we perform LV and myocardial segmentation by regression of pose and shape parameters derived from a statistical shape model. The integrated shape model regularizes predicted segmentations and guarantees realistic shapes. Furthermore, in contrast to semantic segmentation, it allows direct calculation of regional measures such as myocardial thickness. We enforce robustness of shape and pose prediction by simultaneously constructing a segmentation distance map during training. We evaluated the proposed method in a fivefold cross validation on a in-house clinical dataset with 75 subjects containing a total of 1539 delineated short-axis slices covering LV from apex to base, and achieved a correlation of 99% for LV area, 94% for myocardial area, 98% for LV dimensions and 88% for regional wall thicknesses. The method was additionally validated on the LVQuan18 and LVQuan19 public datasets and achieved state-of-the-art results.

Motivation & Objective

• Address the limitation of standard CNNs in medical image segmentation, which often produce unrealistic, disconnected, or missing regions due to lack of explicit shape constraints.
• Improve segmentation robustness and anatomical plausibility by integrating a statistical shape model as a prior through regression of shape coefficients.
• Enable direct, accurate calculation of regional cardiac metrics such as myocardial thickness and regional wall thickness by leveraging the shape model parameters.
• Enhance generalization and robustness of shape and pose prediction by jointly training with semantic segmentation via distance map regression.
• Achieve state-of-the-art performance on public cardiac MRI segmentation challenges (LVQuan18/19) while maintaining anatomical realism.

Proposed method

• Construct a statistical shape model from 2D short-axis cardiac MR images using principal component analysis (PCA) on normalized, landmark-based myocardial contours.
• Represent each LV shape using 12 shape coefficients (b1,p, ..., b12,p) derived from the first 12 principal components, capturing over 99% of shape variation.
• Simultaneously regress pose parameters (θp, cx,p, cy,p) representing rotation and center position to normalize shape for consistent coefficient estimation.
• Perform semantic segmentation by regressing signed distance maps (Dp) for endo- and epicardium, with segmentation derived directly from the distance map using a sigmoid function.
• Train the network end-to-end using a multi-task loss combining shape MSE, pose MSE, and a weighted sum of Dice and MSE on binarized distance maps.
• Apply online data augmentation by perturbing pose (position, orientation) and shape coefficients during training, with corresponding image and distance map re-creation via thin-plate spline and cubic spline interpolation.

Experimental results

Research questions

• RQ1 Can joint regression of shape and pose parameters from a statistical shape model improve segmentation accuracy and anatomical plausibility in cardiac MRI compared to standard semantic segmentation?
• RQ2 Does incorporating segmentation via distance map regression with a combined loss function enhance robustness and reduce disconnected or unrealistic segmentations?
• RQ3 To what extent does online data augmentation of pose and shape parameters improve generalization and performance on unseen data?
• RQ4 Can the predicted shape coefficients directly enable accurate computation of regional cardiac metrics such as myocardial thickness and regional wall thickness?
• RQ5 How does the proposed method compare to state-of-the-art approaches on public cardiac MRI segmentation benchmarks (LVQuan18/19) in terms of segmentation and parameter estimation accuracy?

Key findings

• The method achieved a 99% correlation for left ventricular (LV) area, 94% for myocardial area, 98% for LV dimensions, and 88% for regional wall thicknesses on the in-house dataset of 75 subjects with 1539 short-axis slices.
• On the LVQuan18 challenge dataset, the method achieved a mean absolute error (MAE) of 92 mm² for LV area, 121 mm² for myocardial area, 1.52 mm for LV dimensions, and 1.01 mm for relative wall thickness (RWT), outperforming the winner of the LVQuan18 challenge in LV area and myocardial area.
• On the LVQuan19 challenge dataset, the method achieved an MAE of 134 mm² for LV area, 201 mm² for myocardial area, 2.10 mm for LV dimensions, and 1.78 mm for RWT, with higher correlation (ρ) than the top entries in the challenge for most metrics.
• The addition of both semantic segmentation and data augmentation significantly improved shape coefficient regression, reducing landmark prediction error to 1.44 mm (vs. 2.10 mm without segmentation and 1.85 mm without augmentation).
• The method produced no disconnected or missing regions in segmentations due to the inherent shape model constraint, unlike standard semantic segmentation with Dice loss.
• The method demonstrated superior performance in estimating regional metrics: the correlation for regional wall thickness was 88%, and the MAE for LV dimensions was 1.52 mm on LVQuan18, outperforming the baseline method in [11] by a significant margin.