Temporal Uncertainty Localization to Enable Human-in-the-loop Analysis of Dynamic Contrast-enhanced Cardiac MRI...

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Artificial Intelligence in Medicine

Background:

• Dynamic contrast-enhanced cardiac magnetic resonance imaging (DCE-CMRI) is crucial for assessing myocardial perfusion abnormalities.
• Manual segmentation of DCE-CMRI data is time-consuming and prone to errors, especially with motion artifacts.
• Existing deep neural network (DNN) segmentation methods lack reliable dynamic quality control (dQC).

Approach:

• Developed a novel space-time uncertainty metric for dQC in DNN-based DCE-CMRI segmentation.
• Implemented a human-in-the-loop framework, referring the top 10% most uncertain segmentations for expert refinement.
• Validated the approach on an external dataset to assess its effectiveness.

Key Points:

• The proposed dQC tool successfully identified and flagged segmentations requiring human review.
• Referral of uncertain segmentations significantly improved the Dice score.
• Reduced the rate of failed segmentations from 16.2% to 11.3%.

Conclusions:

• The dQC framework accurately detects poor-quality segmentations in DCE-CMRI datasets.
• Enables efficient DNN-based analysis within a human-in-the-loop pipeline for clinical interpretation.
• Facilitates reliable reporting of dynamic CMRI data.