[Paper Review] A Combined Deep-Learning and Deformable-Model Approach to Fully Automatic Segmentation of the Left Ventricle in Cardiac MRI

M. R. Avendi, Arash Kheradvar+1 more|arXiv (Cornell University)|Dec 25, 2015

Medical Image Segmentation Techniques42 references4 citations

TL;DR

This paper presents a fully automatic left ventricle (LV) segmentation method for cardiac MRI using a hybrid deep learning and deformable model approach. It leverages convolutional networks for LV detection and stacked autoencoders for shape inference, which are integrated into a deformable model to enhance accuracy and robustness, achieving a Dice score of 0.94 and 96.69% good contours on the MICCAI 2009 dataset—surpassing state-of-the-art methods.

ABSTRACT

Segmentation of the left ventricle (LV) from cardiac magnetic resonance imaging (MRI) datasets is an essential step for calculation of clinical indices such as ventricular volume and ejection fraction. In this work, we employ deep learning algorithms combined with deformable models to develop and evaluate a fully automatic segmentation tool for the LV from short-axis cardiac MRI datasets. The method employs deep learning algorithms to learn the segmentation task from the ground true data. Convolutional networks are employed to automatically detect the LV chamber in MRI dataset. Stacked autoencoders are utilized to infer the shape of the LV. The inferred shape is incorporated into deformable models to improve the accuracy and robustness of the segmentation. We validated our method using 45 cardiac MR datasets taken from the MICCAI 2009 LV segmentation challenge and showed that it outperforms the state-of-the art methods. Excellent agreement with the ground truth was achieved. Validation metrics, percentage of good contours, Dice metric, average perpendicular distance and conformity, were computed as 96.69%, 0.94, 1.81mm and 0.86, versus those of 79.2%-95.62%, 0.87-0.9, 1.76-2.97mm and 0.67-0.78, obtained by other methods, respectively.

Motivation & Objective

• To develop a fully automatic, robust, and accurate method for left ventricle segmentation in cardiac MRI without manual intervention.
• To overcome limitations of traditional deformable models and deep learning methods by combining their strengths.
• To improve segmentation accuracy in challenging regions such as basal and apical slices, where intensity inhomogeneity and partial volume effects are prominent.
• To reduce dependency on large, diverse training datasets by incorporating learned shape priors through stacked autoencoders.
• To validate the method on a public benchmark dataset and demonstrate superior performance compared to existing state-of-the-art approaches.

Proposed method

• A two-stage deep learning pipeline is used: first, a convolutional neural network (CNN) detects the left ventricle chamber in short-axis cardiac MRI slices.
• Stacked autoencoders infer the global shape of the LV from the input MRI, learning a compact, hierarchical representation of ventricular morphology.
• The inferred shape is used to initialize a deformable model, which refines the segmentation by optimizing contour smoothness and image fidelity.
• The deformable model integrates shape priors from autoencoders to improve convergence and robustness, especially in low-contrast or noisy regions.
• The method is trained and validated on 45 cardiac MRI datasets from the MICCAI 2009 LV segmentation challenge, with data augmentation to mitigate overfitting.
• The framework is implemented on a CPU platform, with segmentation time under 0.5 seconds per slice, and is designed for potential GPU acceleration.

Experimental results

Research questions

• RQ1 Can a hybrid deep learning and deformable model approach achieve superior segmentation accuracy compared to standalone deep learning or deformable models?
• RQ2 How effectively can stacked autoencoders learn and generalize left ventricular shape priors from limited training data?
• RQ3 To what extent does shape initialization from deep autoencoders improve robustness in challenging apical and basal slices?
• RQ4 Can the integration of learned shape priors reduce sensitivity to intensity inhomogeneity and noise in cardiac MRI?
• RQ5 How does the proposed method compare quantitatively to state-of-the-art methods on a standardized benchmark?

Key findings

• The method achieved a percentage of good contours of 96.69% on the MICCAI 2009 dataset, significantly outperforming previous methods (79.2%–95.62%).
• The Dice similarity coefficient reached 0.94, indicating strong overlap with the ground truth, compared to 0.87–0.90 in other state-of-the-art methods.
• The average perpendicular distance between the segmented and ground truth contours was 1.81 mm, lower than the 1.76–2.97 mm range reported by other methods.
• The conformity metric was 0.86, indicating high structural consistency with the reference contours, surpassing the 0.67–0.78 range of competing approaches.
• The method demonstrated robust performance on challenging basal and apical slices, where intensity variations and partial volume effects are most pronounced.
• Segmentation time was under 0.5 seconds per slice, with the majority of time consumed by the convolutional network and deformable model integration.