Digital breast tomosynthesis (DBT) is being pursued by many medical device manufacturers, because its limited tomographic capability directly addresses a major short-coming of full-field digital mammography (FFDM). Namely, due to the nature of projection imaging, false positives can occur because overlapping normal tissue structures sometimes resemble tumors, and false negatives happen because tumors may be masked by normal tissue - particularly in dense breasts of young women. Many studies on DBT have been performed or ongoing and the general sense is that DBT does help for tumor detection, but it may perform worse than FFDM on microcalcification detection and characterization, which are imaging tasks that can indirectly indicate malignancy. The proposed research aims at optimizing image reconstruction algorithms for DBT. Our preliminary results indicate already, for example, that microcalcification imaging in DBT is limited by the sub-optimality of the image reconstruction algorithm: our initial investigations already yield a factor-of-two gain in microcalcification signal-to-noise ratio. As many of the current iterative image reconstruction (IIR) algorithms employed in DBT were essentially borrowed, with little modification, from nuclear medicine imaging there is clearly room for large gains in image quality for tailoring IIR to DBT. The fact that DBT devices are already undergoing clinical trials with sub-optimal image reconstruction algorithms increases the urgency of the proposed research. The goal of the research is to show that clinical utility of DBT can be enhanced significantly by tailoring IIR to application in DBT. The aims of the proposed research are: (1) Construct a framework for IIR addressing the limited angular scan and fixed-dose trade-off of image noise and number of projections. (2) Design optimization-based approaches robust against projection truncation. (3) Develop data consistency checks and automated correction techniques for DBT projections. (4) Design image quality metrics of DBT images for algorithm optimization. (5) Evaluate algorithm performance with human observers.

R01CA158446

2012-04-01

2017-03-31