Deep Learning-Based Reconstruction for Low-Dose Sparse-View Digital Breast Tomosynthesis

Abstract

Purpose: Digital breast tomosynthesis (DBT) is an emerging breast imaging modality, whichreduces tissue overlaps and improves lesion visibility. Due to the nature of ionizing radiation adopted by DBT and the risk of secondary cancer, motivating efforts to lower imaging dose while maintaining image diagnostic quality. A sparse-view acquisition strategy was employed to reduce radiation dose, though it inevitably degrades image quality. To mitigate this, a Deep Residual U-Net (DRU) model was trained to improve image quality and suppress artifacts. Methods: Based on this framework, two strategies were further proposed to investigate the influence of dose on image quality, one strategy is to use sparsely sampled projections with unchanged dose per projection and another strategy is to use sparsely sampled projections with half dose per projection. Dense-view DBT projections were simulated with anthropomorphic digital breast phantoms using ray tracing at 1-degree interval over 50-degree angular range, yielding 51 projections. Sparse-view datasets were generated under two scenarios: (1) 11 and (2) 7 projections, corresponding to 11/51 and 7/51 of the dense-view DBT dose level. To enable training, sparsely sampled data were linear interpolated to match the dense-view sampling, with the dense-view data as ground truth. Deep Residual U-Net (DRU) was trained on sinogram data with two loss functions, L1 and MSE, to evaluate the stability and generalizability. The predicted projections from Deep Residual U-Net (DRU) were reconstructed using the FBP and compared with ground truth dense-view DBT and linear interpolated sparse-view DBT. Results: Results demonstrated that Deep Residual U-Net (DRU) significantly reduces image noise and artifacts, achieving higher PSNR and SSIM compared to the linear interpolation. Deep Residual U-Net (DRU) performance was stable with both L1 and MSE loss functions, converging rapidly and yielding comparable results. The trained Deep Residual U-Net (DRU)generalized well across different dose levels, consistently improving image quality for both11/51 and 7/51 low-dose sparse-view DBT. Conclusions: This study developed a deep learning model for Low-Dose Sparse-View DBT, our findings indicate that deep learning based projection interpolation can enable substantial dose reduction in DBT while maintaining high image quality, supporting its potential for safer and more effective breast cancer detection. Our model shows that deep learning significantly improves image quality, suppresses image noise, and has excellent robustness.