Cardiac Disease Representation Conditioned by Spatio-temporal Priors in Cine-MRI Sequences Using Generative Embedding...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac cine-MRI is crucial for diagnosing heart conditions but suffers from observer variability and limited quantitative measures.
• Ejection fraction is insufficient for differentiating various cardiac pathologies, failing to utilize full functional information from cine-MRI.

Purpose of the Study:

• To develop an automated image analysis method using a conditional generative adversarial network (CGAN) to identify cardiac disease patterns in cine-MRI.
• To create a topological embedding space for cine-MRI data that captures hidden relationships and serves as a disease marker.

Main Methods:

• A CGAN was trained to map the latent space to cine-MRI data, incorporating left ventricle segmentation and velocity fields for focused pattern recognition.
• The trained network generated embeddings for validation slices by minimizing reconstruction error.
• Classification models were used to evaluate these embeddings as disease markers on 16,000 pathological cine-MRI slices.

Main Results:

• The generative representation achieved an average accuracy of 90.04% and an average F1-score of 89.97% in classifying cardiac pathologies.
• The developed embedding space effectively captured salient cardiac patterns and relationships within the cine-MRI data.

Conclusions:

• The study successfully constructed a topological embedding space from generative representations of cine-MRI data.
• This approach fully exploits hidden relationships in cine-MRI, offering a robust method for representing and classifying cardiac diseases.