Segmentation and quantification of infarction without contrast agents via spatiotemporal generative adversarial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Current clinical methods for myocardial infarction (MI) assessment are time-consuming, prone to errors, and lack reproducibility.
• Accurate segmentation and quantification of MI are vital for timely diagnosis and surgical planning.

Purpose of the Study:

• To introduce a novel, contrast-free, and automated deep learning tool for simultaneous MI segmentation and quantification.
• To address the limitations of existing clinical methods for MI assessment using cine MR images.

Main Methods:

• Development of a deep spatiotemporal adversarial network (DSTGAN) utilizing a conditional generative model.
• Implementation of a multi-level, multi-scale encoder for hierarchical feature extraction of MI abnormalities.
• Integration of cross-task generators and discriminators for enhanced segmentation and quantification accuracy.

Main Results:

• The DSTGAN achieved a pixel classification accuracy of 96.98%.
• The mean absolute error for MI centroid localization was 0.96 mm across 165 clinical subjects.
• Demonstrated stable and automatic performance in segmenting and quantifying MI directly from cine MR images.

Conclusions:

• The proposed DSTGAN method shows significant potential for standardized and reliable MI assessments.
• This contrast-free approach offers a promising alternative for clinical applications in cardiology.
• The study highlights the efficacy of deep learning in improving the accuracy and efficiency of cardiovascular imaging analysis.