A probabilistic deep motion model for unsupervised cardiac shape anomaly assessment.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac shape analysis is crucial for diagnosing pathologies and assessing segmentation quality.
• Manual analysis of large-scale cardiac imaging data is labor-intensive and prone to errors.
• Automated methods are needed to efficiently screen suboptimal segmentations and cardiac morphology alterations.

Purpose of the Study:

• To develop a deep probabilistic model for automatic shape anomaly detection in cardiac imaging.
• To identify suboptimal segmentations and pathologies affecting cardiac morphology.
• To provide an efficient tool for large-scale cardiac imaging pipelines.

Main Methods:

• A deep recurrent encoder-decoder network was used to model spatio-temporal dependencies in cardiac shape sequences.
• A predictive mixture distribution with Gaussian (inlier) and uniform (outlier) components was employed.
• A Gibbs sampling Expectation-Maximisation (EM) algorithm was utilized for anomaly scoring and parameter estimation.

Main Results:

• The model successfully detected shape anomalies in large datasets (UK Biobank, M&Ms), correlating with segmentation quality and pathologies.
• Anomalies in UK Biobank data were linked to poor segmentation, while M&Ms data revealed links to ventricular pathologies.
• The model demonstrated significant improvement in predicted cardiac shape sequences over input sequences.

Conclusions:

• The proposed deep probabilistic model is effective for automatic shape anomaly detection in cardiac imaging.
• This method can efficiently screen large datasets for segmentation issues and disease-related morphological changes.
• The model offers a valuable tool for enhancing the analysis of cardiac imaging data in clinical and research settings.