[Paper Review] Automated cardiovascular magnetic resonance image analysis with fully convolutional networks

Wenjia Bai, Matthew Sinclair+21 more|arXiv (Cornell University)|Oct 25, 2017

Cardiac Imaging and Diagnostics24 references4 citations

TL;DR

This paper proposes a fully convolutional network (FCN)-based method for automated segmentation of cardiac structures in short-axis and long-axis cardiovascular magnetic resonance (CMR) images. Trained on 4,875 UK Biobank subjects with 93,500 pixel-wise annotations, the model achieves human-level accuracy in measuring left and right ventricular volumes, ejection fraction, and myocardial mass, with Dice scores of 0.94 for left ventricular cavity and mean absolute differences of 6.1 mL for end-diastolic volume compared to manual tracing.

ABSTRACT

Cardiovascular magnetic resonance (CMR) imaging is a standard imaging modality for assessing cardiovascular diseases (CVDs), the leading cause of death globally. CMR enables accurate quantification of the cardiac chamber volume, ejection fraction and myocardial mass, providing information for diagnosis and monitoring of CVDs. However, for years, clinicians have been relying on manual approaches for CMR image analysis, which is time consuming and prone to subjective errors. It is a major clinical challenge to automatically derive quantitative and clinically relevant information from CMR images. Deep neural networks have shown a great potential in image pattern recognition and segmentation for a variety of tasks. Here we demonstrate an automated analysis method for CMR images, which is based on a fully convolutional network (FCN). The network is trained and evaluated on a large-scale dataset from the UK Biobank, consisting of 4,875 subjects with 93,500 pixelwise annotated images. The performance of the method has been evaluated using a number of technical metrics, including the Dice metric, mean contour distance and Hausdorff distance, as well as clinically relevant measures, including left ventricle (LV) end-diastolic volume (LVEDV) and end-systolic volume (LVESV), LV mass (LVM); right ventricle (RV) end-diastolic volume (RVEDV) and end-systolic volume (RVESV). By combining FCN with a large-scale annotated dataset, the proposed automated method achieves a high performance on par with human experts in segmenting the LV and RV on short-axis CMR images and the left atrium (LA) and right atrium (RA) on long-axis CMR images.

Motivation & Objective

• To address the clinical challenge of time-consuming and subjective manual segmentation of CMR images by developing an automated, objective alternative.
• To enable large-scale population imaging studies, such as the UK Biobank, which require high-throughput, accurate analysis of tens of thousands of CMR scans.
• To demonstrate that deep learning models can match or exceed human expert performance in segmenting cardiac structures across multiple cardiac views.
• To establish a benchmark for automated CMR analysis using a large-scale, publicly available, manually annotated dataset.

Proposed method

• A fully convolutional network (FCN) is trained end-to-end to perform pixel-wise segmentation of cardiac structures on short-axis and long-axis CMR images.
• The model uses multi-scale feature fusion through skip connections to improve localization accuracy in segmentation.
• Training is performed on a large-scale dataset of 4,875 UK Biobank subjects with 93,500 manually annotated images, covering left and right ventricles, left and right atria.
• The network is optimized using a combination of cross-entropy and dice loss to improve boundary detection and segmentation consistency.
• Data pre-processing includes intensity normalization, image resizing, and augmentation to improve generalization and robustness.
• Post-processing includes morphological operations and contour refinement to ensure anatomically plausible segmentation outputs.

Experimental results

Research questions

• RQ1 Can a fully convolutional network achieve human-level performance in segmenting cardiac structures across short-axis and long-axis CMR images?
• RQ2 How does the performance of the automated method compare to inter-observer variability among human experts?
• RQ3 To what extent can deep learning reduce the time and subjectivity associated with manual CMR image analysis?
• RQ4 Can a large-scale, publicly available annotated dataset support the development of robust, generalizable CMR analysis models?

Key findings

• On a test set of 600 short-axis subjects, the method achieved a mean Dice score of 0.94 for left ventricular cavity, 0.88 for left ventricular myocardium, and 0.90 for right ventricular cavity.
• The mean absolute difference between automated and manual measurements was 6.1 mL for left ventricular end-diastolic volume, 5.3 mL for end-systolic volume, and 6.9 g for left ventricular mass.
• For right ventricular volumes, the mean absolute difference was 8.5 mL for end-diastolic volume and 7.2 mL for end-systolic volume.
• On long-axis images, the model achieved a Dice score of 0.93 for the left atrium (2-chamber view), 0.95 for the left atrium (4-chamber view), and 0.96 for the right atrium (4-chamber view).
• The performance of the automated method was comparable to human inter-observer variability, indicating clinical readiness.
• The method enables automated, fast, and objective quantification of clinically relevant measures such as ejection fraction and stroke volume within seconds per study.