[Paper Review] Human-level CMR image analysis with deep fully convolutional networks.

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

Medical Image Segmentation Techniques33 citations

TL;DR

This paper presents a deep fully convolutional network (FCN) for automated, human-level analysis of cardiovascular magnetic resonance (CMR) images. Trained on a large-scale dataset of 4,875 subjects with 93,500 pixelwise annotated images, the model achieves performance on par with expert radiologists in quantifying cardiac chamber volume, ejection fraction, and myocardial mass.

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 a wealth of information for sensitive and specific 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 dataset of unprecedented size, consisting of 4,875 subjects with 93,500 pixelwise annotated images, which is by far the largest annotated CMR dataset. By combining FCN with a large-scale annotated dataset, we show for the first time that an automated method achieves a performance on par with human experts in analysing CMR images and deriving clinical measures. We anticipate this to be a starting point for automated and comprehensive CMR analysis with human-level performance, facilitated by machine learning. It is an important advance on the pathway towards computer-assisted CVD assessment.

Motivation & Objective

• To address the clinical challenge of time-consuming and subjective manual analysis of CMR images.
• To develop an automated method that achieves diagnostic accuracy comparable to human experts in CMR image analysis.
• To leverage deep learning for robust, pixelwise segmentation of cardiac structures in CMR scans.
• To demonstrate the feasibility of large-scale, fully convolutional networks in medical image analysis with clinical relevance.
• To establish a foundation for computer-assisted, comprehensive CMR assessment using machine learning.

Proposed method

• A fully convolutional network (FCN) is employed to perform end-to-end, pixelwise segmentation of cardiac structures in CMR images.
• The model is trained on a large-scale dataset of 4,875 subjects, comprising 93,500 pixelwise annotated CMR images—the largest such dataset to date.
• The FCN architecture enables dense prediction across the entire image, preserving spatial resolution for accurate anatomical localization.
• The network is optimized using a loss function that supports end-to-end learning for segmentation tasks.
• The method is evaluated using standard metrics for cardiac segmentation and clinical quantification.
• Performance is benchmarked against expert-level human readers to validate clinical relevance.

Experimental results

Research questions

• RQ1 Can a deep fully convolutional network achieve human-level performance in CMR image analysis?
• RQ2 Does a large-scale, pixelwise annotated dataset significantly improve the accuracy of automated CMR analysis?
• RQ3 Can an FCN-based method reliably derive clinically relevant measures such as ejection fraction and myocardial mass from CMR images?
• RQ4 How does the performance of the automated method compare to expert human readers in quantitative cardiac assessment?
• RQ5 Can deep learning models be effectively scaled to clinical-grade CMR image analysis with high reliability?

Key findings

• The proposed FCN model achieves performance on par with human experts in analyzing CMR images and deriving key clinical metrics.
• The model demonstrates high accuracy in segmenting left ventricular endocardial and epicardial borders across the entire cardiac cycle.
• The use of a large-scale dataset of 93,500 pixelwise annotated images significantly enhances model generalization and robustness.
• The method enables automated, reproducible quantification of cardiac chamber volume, ejection fraction, and myocardial mass.
• The results represent a milestone in automated CMR analysis, marking the first time a deep learning model reaches human-level performance in this domain.
• The approach paves the way for scalable, computer-assisted cardiovascular disease assessment in clinical settings.