Deep learning based tomosynthesis denoising: a bias investigation across different breast types

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low X-ray dose in digital breast tomosynthesis (DBT) leads to high noise, challenging image reconstruction.
• Deep learning (DL) offers potential for noise reduction but risks bias if training data is unrepresentative.
• Thorough evaluation of DL denoising algorithms is crucial for medical application safety and fairness.

Purpose of the Study:

• To develop and evaluate a deep learning-based denoising algorithm for DBT.
• To investigate potential biases of the algorithm concerning breast density, thickness, and noise levels.

Main Methods:

• Physics-driven data augmentation was used to create low-dose images from full-field digital mammography.
• An encoder-decoder network was trained using a novel rectified linear unit (ReLU)-loss function tailored for mammographic denoising.
• Algorithm performance and bias were assessed using both clinical and simulated data, including varied X-ray dose distributions.

Main Results:

• Denoising performance demonstrated a direct correlation with noise level.
• While simulated data indicated bias towards specific breast groups, clinical data showed the algorithm performed equally well across different breast types based on structural similarity index.
• The algorithm effectively reduced noise in DBT projection images.

Conclusions:

• A robust deep learning denoising algorithm for DBT was developed.
• Extensive testing revealed the algorithm's strengths and weaknesses, particularly its unbiased performance on diverse breast types in clinical data.