The development of mathematical observers for optimization in breast imaging

Breast cancer screening in women aims to reduce cancer mortality rate by means of early detection and treatment. The screening is typically  performed using a two-dimensional (2D) full-field digital mammography (FFDM) system, where the breast structure is captured in a single 2D projection. However the representation of three-dimensional (3D) breast tissue in 2D images causes a major drawback of this technique: normal tissue could mask a malignancy or overlapping normal tissue could appear suspicious. To reduce this drawback the digital breast tomosynthesis (DBT) technique was introduced. In DBT, series of projections are acquired over a narrow angular range around the breast. These projections are then reconstructed into a pseudo-3D volume, which allows for cross-sectional visualization of the breast and partly solves the tissue overlapping  found in 2D FFDM.

The major radiological breast cancer signs are low contrast large objects (masses) and tiny high-contrast objects (microcalcifications) in a background of sometimes very complicated breast tissue patterns in the 2D mammogram or the DBT reconstructed volume. Breast cancer detection is a challenging task for both the imaging system and the radiologist. For DBT and FFDM the diagnostic information is inversely correlated with the X-ray radiation exposure needed for the medical examination. Therefore high image quality is aimed for at a mean glandular dose to the women that is as low as reasonably achievable. In order to fulfill this requirement the mammography system performance is regularly evaluated in terms of image quality and dose. In order to balance between diagnostic information and X-ray exposure risk, research studies are conducted to optimize the X-ray systems. The standard means of image quality evaluation for these purposes typically involve a test object with critical targets and human observer reading studies. While these studies are the gold standard, they usually require a substantial amount of time, especially if there are many parameters involved, and might turn out as a bottleneck for certain optimization studies. This thesis focuses on the development of an automated method for image quality evaluation using model observers to replace human observers for quality assurance and optimization in breast imaging.

The statistical model observers and specifically the channelized Hotelling observer (CHO) are mathematical operations based on the statistical properties of the pixel values in the images. A possible application, after specific adjustments of the algorithms, is related to the estimation of X-ray system performance in terms of detectability of certain lesions in images. A recently developed 3D structured phantom in our research group showed that it can be successfully used to evaluate image quality on different DBT and FFDM systems under different scanning conditions using human observers. We hypothesized that the phantom evaluation can be achieved using a tuned CHO algorithm that predicts the human observer scores. This would allow for quicker and more reliable observer performance estimations.

The first step was to study and evaluate the effect of all different components contributing to the CHO performance estimation for low contrast lesions representing non-spiculated masses. The successfully developed CHO model had high correlation with the human observers using a Siemens DBT system. The CHO was further applied to other DBT vendors, which showed poor correlation with human results. In order to achieve good generalization to a wider range of DBT systems, the channel mechanism was improved to select the same frequency range regardless of the reconstruction pixel size of the different DBT systems and the first comparative studies with the CHO on all commercially available systems was performed. The next step focused on the small high-contrast objects, namely calcification clusters. With these targets, largely differing from mass lesions, a more innovative CHO approach was required. The calcification clusters in the phantom consist of 5 calcification particles forming the target. A two-layer CHO algorithm was developed for the cluster detection in DBT and FFDM. First and for the Siemens DBT system, the particle locations were found by scanning areas around the expected locations, and this was followed by a classification step, where the separate particle test statistics were combined into a single cluster test statistics. Exactly the same algorithm was also tested for different DBT vendors and showed good correlation without a requirement for additional tuning. With a CHO validated to work in different scanning conditions and for multiple DBT vendors, the CHO was tested on different 3D structured DBT test images. The virtual clinical trial image dataset obtained from the OPTIMAM simulation framework consisted of 2D FFDM images and DBT images with simulated mass lesions. The already developed CHO algorithm was redesigned to work also on the 2D images and the CHO observer performance for the DBT systems was further improved by introducing a volumetric CHO. This update of the CHO, required to make the CHO work on these different backgrounds showed that the previously developed algorithm did not generalize well enough. We hypothesized then that the channelization algorithm (feature extraction algorithm) for mass-like lesions can be improved using deep learning. Thus two deep learning model observers were developed in parallel to the standard CHO. Both deep learning methods were trained on 4-AFC examples from humans. This resulted in better generalization across a wider range of DBT vendors and conditions.

In summary,  this thesis investigated the feasibility of a channelized Hotelling observer to estimate the human observer reading results of image quality evaluation tasks. Two distinct practical tools emerged from the work summarized in this PhD project: a deep learning CHO for mass-like lesion detection and a two-layer CHO for calcification cluster detection. It is our hope that these tools will be used routinely by medical physicists and researchers for image quality evaluation studies. The software is fully prepared for use by many users and a specific teaching course has been developed. We hope that it may help to give DBT the role it deserves in the breast cancer screening, and ultimately bring mortality down.