Denoising digital breast tomosynthesis projections using deep learning with synthetic data as training set.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks (DNNs) for image denoising require extensive datasets, which are difficult to obtain for digital breast tomosynthesis (DBT) due to varying radiation doses.
• The lack of sufficient real-world training data hinders the development of effective DNN-based denoising techniques for DBT.

Purpose of the Study:

• To investigate the efficacy of using synthetic data for training DNNs to denoise DBT projections.
• To address the challenge of limited real-world data availability in DBT image processing.

Main Methods:

• Generated a synthetic dataset of DBT images with varying noise levels using software (OpenVCT and photographic synthesis with Poisson-Gaussian noise models).
• Trained DNN denoising models on the synthetic dataset and evaluated their performance on real DBT data using quantitative (PSNR, SSIM) and qualitative measures.
• Employed t-SNE for visualizing and comparing the sample spaces of synthetic and real datasets.

Main Results:

• DNN models trained with synthetic data successfully denoise real DBT images.
• Quantitative results were competitive with traditional denoising methods.
• Qualitative analysis revealed superior noise filtering and detail preservation compared to traditional methods.

Conclusions:

• Synthetic data is a viable and effective solution for training DNNs for DBT image denoising.
• The key is ensuring the synthesized noise resides within the same sample space as the target real images.
• This approach overcomes the data acquisition bottleneck for developing advanced DBT denoising tools.