Convolutional neural network -based phantom image scoring for mammography quality control.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology Quality Assurance

Background:

• Mammography quality control (QC) relies on time-consuming visual phantom image evaluation.
• Consistent scoring of mammography phantom images is crucial for device longevity.
• Convolutional neural networks (CNNs) show high accuracy in image classification tasks.

Purpose of the Study:

• To automate mammography QC phantom scoring using CNN models.
• To train CNNs to replicate human reviewer performance in phantom image analysis.

Main Methods:

• Trained eight CNN variations (3-10 layers) to detect targets in American College of Radiology (ACR) accreditation phantom images.
• Compared CNN results with human scoring on regular, degraded, and improved QC phantom images from eight devices.
• Validated CNN performance against ground truth scores from four human reviewers on daily QC and varying dose images.

Main Results:

• An optimal CNN network depth was identified, with six convolutional layers achieving the highest accuracy (95%).
• CNN performance showed the highest deviation from human reviewers at the lowest dose levels.
• CNN results closely matched human reviews, with minor differences noted for the smallest masses.

Conclusions:

• A CNN-based system can automate mammography QC phantom scoring effectively.
• The automated system demonstrates good agreement with human reviewers.
• This technology can significantly benefit mammography quality control processes.