Towards reliable cardiac image segmentation: Assessing image-level and pixel-level segmentation quality via...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Diseases

Background:

• Deep learning models excel in cardiac image segmentation but struggle in clinical deployment due to domain shifts and artifacts.
• Ensuring reliable computer-aided diagnosis requires methods to identify and flag poor-quality segmentations.

Purpose of the Study:

• To develop a robust quality control framework for identifying failure segmentations in cardiac imaging.
• To provide physicians with reliable information to avoid diagnostic errors stemming from inaccurate AI segmentations.

Main Methods:

• A reference-based framework assessing image-level (per-class Dice) and pixel-level (pixel-wise correct map) quality.
• Generating references by reconstructing input images from segmentations using a self-reflective generator.
• Employing a difference investigator with a semantic class-aware compactness constraint to analyze reconstruction inconsistencies.

Main Results:

• The proposed method effectively identifies segmentation failures across various quality levels (low, medium, high).
• Experiments on ACDC and MSCMR datasets demonstrate the method's robustness and ability to capture segmentation errors.
• The framework successfully distinguishes between high-quality and poor-quality segmentations.

Conclusions:

• The developed quality control method enhances the reliability of deep learning-based cardiac image segmentation in clinical practice.
• This approach supports informed clinical decision-making by highlighting potentially erroneous segmentations.
• The self-reflective reference generation and difference investigation offer a novel way to assess segmentation quality.