Automated cardiovascular magnetic resonance image analysis with fully convolutional networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiovascular resonance (CMR) imaging is crucial for diagnosing cardiovascular diseases (CVDs).
• Manual CMR image analysis is time-consuming and prone to subjective errors.
• Automating quantitative analysis of CMR images presents a significant clinical challenge.

Purpose of the Study:

• To develop and evaluate an automated method for analyzing CMR images using deep neural networks.
• To enable accurate and efficient quantification of cardiac structures from CMR data.
• To overcome the limitations of manual CMR image analysis.

Main Methods:

• A fully convolutional network (FCN) was developed for automated CMR image analysis.
• The FCN was trained and validated on a large-scale UK Biobank dataset (4,875 subjects, 93,500 images).
• Performance was assessed using technical metrics (Dice, contour distance) and clinical measures (volumes, mass).

Main Results:

• The automated method achieved high performance in segmenting left ventricle (LV), right ventricle (RV), left atrium (LA), and right atrium (RA).
• Achieved Dice metrics of 0.94 (LV cavity), 0.88 (LV myocardium), and 0.90 (RV cavity) on short-axis images.
• Clinical measures showed minimal differences compared to manual analysis, comparable to human inter-observer variability.

Conclusions:

• An automated FCN-based method provides accurate CMR image analysis.
• The automated method's performance is on par with human experts.
• This approach offers a potential solution for efficient and reliable CVD diagnosis and monitoring.