Digital Breast Tomosynthesis (DBT) has been shown to significantly improve the detection and characterization of soft-tissue lesions and reduce false positive recalls in breast cancer screening. However, DBT is still at its early stage of clinical use and continued improvement of the system design and reconstruction methods are crucial to fully exploit its potential. Noise and resolution are major factors in optimization of an imaging system. The noise in DBT is much higher than that in digital mammograms (DMs) because the multiple low-dose projections increase the total detector noise. The oblique incidence to the breast and the detector at large-angle projections further aggravates the noise problem and reduces spatial resolution. Synthetic mammograms cannot resolve these problems because they are generated from the DBT. It is known from CT that iterative reconstruction (IR) with properly designed regularizer can significantly reduce noise. However, IR for CT generally does not consider spatial blur and noise correlation/aliasing. Modeling these factors has recently started in CBCT that uses flat panel detectors. Model-based IR (MBIR) technology has not been developed for DBT. DBT is a limited-angle tomography, which, coupled with the very different target signals that are signs of breast cancer (microcalcifications, spiculated/ill-defined masses and distortions) than those in CT or CBCT, makes it much more challenging to develop MBIR for DBT. The goal of the proposed project is to develop MBIR for DBT by accurate physics and statistics modeling of the imaging system to improve the image quality of DBT. We will develop accelerated reconstruction algorithms for these models to facilitate both research and eventual translation to clinical use of such methods.
Our specific aims are: (SA1) prepare three data sets for development of the MBIR method (simulated DBT projection data, DBT projections of physical phantoms, and human subject DBT projections), and study the impacts of various image degrading factors on the reconstructed DBT; (SA2) develop MBIR by optimizing the design of the objective function and the iterative algorithm using the three types of data obtained in (SA1) and a four-tier approach; and (SA3) validate the developed MBIR method by comparison with current reconstruction techniques in terms of the detection accuracy of target signals by radiologists (ROC study) and by computer-aided detection (CAD) systems in human subject DBT images. This project brings together two research teams with complementary expertise, one in imaging physics, image analysis and lesion detection in DBT, the other in statistical iterative reconstruction for CT/SPECT/ PET/MRI, to tackle this limited-angle reconstruction problem. If successful, DBT reconstructed with the new MBIR method is expected to improve the efficacy of early breast cancer detection and diagnosis and reduce dose. Reducing dose and noise will also facilitate the optimization of overall DBT system design, and development of advanced DBT techniques such as dual-energy contrast-enhanced DBT or dynamic contrast-enhanced DBT, which may be cost-effective alternatives to breast MRI for cancer diagnosis.
Digital breast tomosynthesis (DBT) is a promising new modality for screening and diagnosis of breast cancer. Model-based iterative reconstruction with accurate physics modeling of the DBT system will improve the resolution and reduce noise of the reconstructed DBT. The proposed project therefore has the potential to improve the detection and diagnosis of early stage breast cancer and reduce radiation risk using DBT. Reducing dose will also encourage compliance with screening recommendation, facilitate optimization of DBT system design and the development of advanced DBT techniques such as dual-energy contrast-enhanced DBT or dynamic contrast-enhanced DBT that may be cost-effective alternatives to breast MR for cancer diagnosis. In addition, this research will bring significant technological advancement and valuable knowledge to limited-angle CT reconstruction or tomosynthesis in general, which may be a low-dose alternative to CT for many imaging procedures.
|Samala, Ravi K; Chan, Heang-Ping; Hadjiiski, Lubomir M et al. (2018) Evolutionary pruning of transfer learned deep convolutional neural network for breast cancer diagnosis in digital breast tomosynthesis. Phys Med Biol 63:095005|