Automatic quality control methods for mammography and image quality of digital breast tomosynthesis systems

Breast cancer is the most common cancer among women worldwide, and it is often detected using mammography. Mammography is a two-dimensional imaging modality dedicated to breast imaging, which utilizes x-rays for image formation. To overcome limitations of mammography, digital breast tomosynthesis has also been developed. Mammography has a high image quality requirement, which necessitates systematic quality control. The aims of this dissertation were to develop automatic quality control methods for mammography and determine technical and subjective image quality of five digital breast tomosynthesis systems. Daily image quality control in HUS Diagnostic Center is performed using the American College of Radiology mammography accreditation phantom which contains fibers, microcalcification groups and masses. Phantom images are usually evaluated visually. To overcome limitations of visual evaluation, the phantom scoring task was automatized using wavelet transform and convolutional neural networks. To study technical image quality of five digital breast tomosynthesis systems, several image quality parameters were measured using phantoms. The subjective image quality of these digital breast tomosynthesis systems was studied using radiologists’ reviews of phantom images containing filament, microcalcification group and low-contrast detail targets. Both automated image quality analysis software resulted in similar target detection levels with visual evaluations made by the reviewers. The convolutional neural network-based scoring algorithm was slightly more accurate than the wavelet transform-based algorithm. In technical image quality evaluations of digital breast tomosynthesis systems notable differences between systems were observed in artefact spread in the z-direction, in-plane resolution, and noise power spectra of reconstructed planes. Notable in-plane resolution anisotropy was observed between x- and y-directions for most of the systems. The geometric accuracy was within 0.5% for every system. In subjective image quality reviews at similar dose levels, only one system differed significantly from all other systems in microcalcification scores. However, when using site-specific protocols, significant differences were found between the systems for all target scores. Moreover, a strong correlation was found between the reviewer scores and radiation doses in acquisitions. In conclusion, the developed automatic mammography quality control phantom scoring software provides an efficient method for monitoring the function and performance of digital mammography systems. This benefits especially a multi-unit center with systems from several vendors and different generations. The studied digital breast tomosynthesis systems have different technical and image reconstruction solutions, which lead to notable differences in technical image quality. The subjective image quality between digital breast tomosynthesis systems was found to be mostly related to dose level.