Advances in Image Reconstruction for Digital Breast Tomosynthesis

Abstract

Digital breast tomosynthesis (DBT) is an important imaging modality for breast cancer screening and diagnosis. It acquires a sequence of projection views within a limited angle and provides quasi-three-dimensional images of the breasts, allowing for improved lesion visualization and reduced false positives compared with two-dimensional mammography. Despite its advantages, DBT suffers from noise and blur problems that can compromise image quality and reduce its sensitivity in detecting subtle signs of breast cancer such as microcalcifications (MCs). The primary objective of this thesis is to push the state-of-the-art of DBT imaging by developing advanced DBT image reconstruction and processing methods. By reducing image noise, enhancing spatial resolution, optimizing reconstruction methods and evaluating them based on clinical tasks, our ultimate goal is to make DBT an even more effective tool for breast cancer screening and diagnosis. In this thesis, we first developed a deep convolutional neural network (DCNN) for denoising reconstructed DBT images. We trained the DCNN using a weighted combination of mean squared error loss and the adversarial loss based on generative adversarial network (GAN), and therefore called it DNGAN. The DNGAN improved the contrast-to-noise ratio, detectability index, and human observer detection sensitivity of the MCs in DBT images of breast simulating phantoms. Promising denoising results were also observed on a small test set of human subject DBTs. Then, we introduced a model-based DCNN-regularized reconstruction (MDR) method for DBT. It combined a model-based iterative reconstruction method with the DNGAN denoiser. To facilitate task-based image quality assessment, we also proposed two DCNN tools: CNN-NE for noise estimation, and CNN-MC as a model observer for MC cluster detectability measure. We demonstrated the effectiveness of CNN-NE and CNN-MC using phantom DBTs. The MDR method achieved low noise and the highest detection rankings on a test set of human subject DBTs. Finally, we presented our work on modeling the x-ray source motion blur of the DBT imaging system. We derived an analytical in-plane source blur kernel for DBT images based on imaging geometry and showed that it could be approximated by a shift-invariant kernel over the DBT slice at a given height above the detector. We proposed a post-processing image deblurring method with a generative diffusion model as an image prior and successfully enhanced spatial resolution of the reconstructed DBT images.