Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Cardiovascular diagnostics

Background:

• Accurate delineation of cardiac structures in cardiac magnetic resonance imaging (CMR) is crucial for clinical diagnosis.
• Automating these segmentation and classification tasks has been a long-standing research goal in cardiovascular imaging.

Purpose of the Study:

• To introduce the Automatic Cardiac Diagnosis Challenge (ACDC) dataset, the largest publicly available annotated dataset for cardiac MRI assessment.
• To evaluate the performance of state-of-the-art deep learning methods for cardiac MRI segmentation and pathology classification.

Main Methods:

• Development and release of the ACDC dataset, comprising 150 multi-equipment CMR recordings with expert annotations.
• Benchmarking deep learning models from nine research groups for segmentation and four for classification tasks within the 2017 MICCAI-ACDC challenge framework.

Main Results:

• The best deep learning methods achieved a 0.97 correlation score for automatic extraction of clinical indices and 0.96 accuracy for automatic diagnosis, closely matching expert performance.
• Identified specific scenarios where deep learning methods still face challenges in cardiac MRI analysis.

Conclusions:

• Deep learning models demonstrate high potential for fully automatic and accurate analysis of cardiac MRI.
• The ACDC dataset and challenge results provide a valuable resource for advancing automated cardiovascular diagnostics and highlight areas for future research.