Task-based assessment of digital mammography microcalcification detection with deep learning denoising algorithmss...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Breast Cancer Diagnostics

Background:

• Mammography is crucial for breast cancer detection, identifying microcalcifications and lesions.
• Reduced radiation exposure in mammography can degrade image quality, impacting biomarker visibility.
• Denoising algorithms, particularly those using convolutional neural networks (CNNs), show promise in mitigating this quality loss.

Purpose of the Study:

• To evaluate the effectiveness of a deep learning-based denoising method in improving the detection of small microcalcifications in half-dose mammograms.
• To assess if postprocessing with a CNN denoiser enhances the visibility of microcalcifications, a key indicator in mammography.

Main Methods:

• An existing CNN denoiser model was applied to physical phantom mammograms and simulated breast phantoms.
• Human reader studies using 4-alternative forced choice (4-AFC) experiments were conducted to compare image quality and detection performance.
• Performance was measured by the proportion of correct responses in signal detection tasks.

Main Results:

• No significant improvement in detecting small microcalcifications was observed in denoised half-dose mammograms from physical phantom studies.
• A notable increase in 4-AFC scores was found in a Monte Carlo simulation study using denoised half-dose images.
• This improvement in simulation occurred when the neural network was trained on a dataset with a balanced mix of signal-present and signal-absent regions.

Conclusions:

• Deep learning denoising may not inherently improve microcalcification detection in all mammographic scenarios.
• Enriching training datasets with signal-present regions of interest (ROIs) could enhance the performance of these algorithms.
• This enhancement is particularly relevant for detecting clusters of small microcalcifications ($$).