Consistent explainable image quality assessment for medical imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate medical image quality assessment is vital to prevent misdiagnosis.
• Manual quality assessment is time-consuming and prone to errors in large datasets.
• Automated methods lack explainability and robust quantification.

Purpose of the Study:

• To propose an explainable automated system for medical image quality assessment.
• To validate the system in foreign object detection (Chest X-Rays) and cardiac MRI.
• To quantify the reliability and explainability of the proposed methodology.

Main Methods:

• Utilized NormGrad, an algorithm comparing classifier saliency maps with baseline detectors.
• Introduced the Difference of Means (DoM) metric to evaluate saliency detector consistency.
• Validated the system on Object-CXR and LVOT Cardiac MRI datasets.

Main Results:

• NormGrad significantly outperformed other saliency detectors with high Pointing Game scores (0.863 for Object-CXR, 0.778 for LVOT).
• The system demonstrated high consistency across different models, with low DoM scores (0.001 for Object-CXR, 0.005 for LVOT).
• Publicly available code and experiments facilitate reproducibility.

Conclusions:

• The NormGrad-powered system enhances the reliability of automated medical image quality evaluation.
• The DoM metric provides a novel approach to assess saliency detector consistency.
• The system shows promise for clinical adoption in medical imaging quality control.