Abstract
Breast cancer continues to be the second most frequently diagnosed cancer among US women, coming second only to skin cancers. The current standard for breast cancer screening is x-ray mammography (XRM), however XRM sensitivity and specificity decrease with increasingly radiodense breasts. Imaging modalities such as ultrasound (US) and magnetic resonance (MRI) are utilized as effective supplements for mammography, however, used as stand-alone screening modalities US and MRI suffer from high false positive rates. This dissertation evaluates two unique nuclear medicine (NM) based imaging systems developed with the goal of improving the detection and characterization of breast abnormalities.
The first system is a dual modality tomosynthesis (DMT) imaging system. This system includes two integrated components, an x-ray digital breast tomosynthesis (DBT) system and a novel NM system for molecular breast tomosynthesis (MBT). DBT and MBT are performed on a shared gantry, allowing for a spatially co-registered hybrid set of 3D x-ray transmission and gamma ray emission images. A human study was performed to assess the value of adding functional 3D MBT images to the current standard of combined 3D DBT plus 2D XRM. In this study 94 subjects, all scheduled for biopsies, were imaged. Of these, actual biopsies and full image sets were obtained for 75. A total of 83 lesions were biopsied with 21 found to be malignant and 62 benign. A NM radiologist first interpreted the MBT images alone localizing any findings and rating them on a linear suspicion scale from 1-5, with 5 being definitely malignant. The findings from the NM radiologist were then used as a consult report for a breast radiologist who interpreted the images for each case (both breasts independently) in the following sequence: 1) DBT alone, 2) DBT + XRM, 3) DMT (=DBT+MBT) + XRM, 4) DMT + XRM + consult report. Updated suspicion scores were provided as each additional type of information was added. Reader interpretation results for each modality were then compared in terms of the area under the receiver operating characteristic (ROC) curve (AUC), using the biopsy results as ground truth. Compared to the reference imaging standard of DBT+XRM (AUC = 0.74), all modalities, DMT + XRM (AUC = 0.93), MBT (AUC = 0.90), and DBT (AUC = 0.58), are all shown to be significantly different. Both DMT +XRM (difference in AUC – 0.19, 95% confidence limits – 0.075 to 0.302, p-value 0.0011) and MBT alone (difference – 0.16, 95% confidence limits – 0.028 to 0.293, p-value 0.0177) showed a significant improvement in diagnostic accuracy, while DBT alone (difference – -0.17, 95% confidence limits – 0.283 to 0.048, p-value 0.0056) showed a significant decrease in diagnostic accuracy using 0.05 as an indicator for significance.
To allow for timely evaluation and optimization of system and protocol improvements to the DMT system a computational model observer, based on the channelized Hotelling observer (CHO), is under development. Two versions of the CHO will be developed; one for DBT and one for MBT. To get the large image sets necessary to generate the ensemble covariance matrices required to inform the model observer regarding the statistical noise properties of each imaging task, sets of synthetic DBT and MBT images were created. The synthetic images were created using principle component analysis (PCA) and a subset of the available human DMT data sets, resulting in a set of eigenvectors, which we call eigenDMT images. A variety of synthetic training images were formed via weighted sums of the eigenDMT images plus a mean (averaged over eigenDMT images) breast image. We demonstrated that synthetic images generated from the eigenimage basis set using weighting factors selected for similarity to any particular human image matched that image voxel-by-voxel with a mean square error analysis showing a negligible difference of 4e-8. Synthetic images generated using randomly selected weighting factors were compared to those from the original human DMT images by comparing their power spectral densities. The metric used for quantifying the similarity was the power-law exponent obtained by fitting the power spectral density estimates with a power law of form P(f) = A/f, where f is radial spatial frequency. A two-tailed paired Student’s t-test, with an assumed statistical significance level of p<0.05, was then used to determine if the -values differ significantly or not. The -values of the synthetic DBT and MBT images were shown to not differ significantly from those of the human images. The synthetic images are thus thought to be suitable to be used as a viable and realistic training set for the model observer under development.
A new low-profile (LP) gamma camera, for use in the DMT system, was designed with the goals of increasing the field of view (FOV) and overall imaging performance compared to the DMT gamma camera used to that point. Like the previous camera, the new camera has a NaI(Tl) pixelated scintillation crystal and uses position sensitive photomultiplier tubes (PSPMTs) with a custom-designed electronic readout. Compared to the original camera a higher sensitivity collimator was designed and fabricated. The performance of the two cameras was compared using the following metrics: energy resolution, sensitivity, intrinsic and extrinsic spatial resolution. A breast phantom with simulated lesions was created to compare the signal-to-noise ratio (SNR) of the two cameras under realistic conditions. An improvement in energy resolution was seen (LP camera – 10.8%, original camera – 13.5%). The intrinsic spatial resolution was measured to be similar between cameras; ~2.3 mm FWHM. The LP camera demonstrated a 1.7x increase in sensitivity relative to the original camera. The SNR experiment showed that the lesion SNR was a minimum of 1.7x higher for the LP camera.
The second imaging system is a dedicated breast ring PET (BRPET) scanner. The scanner contains a single ring of 12 small detector modules that surround the pendant breast, while the patient lies prone on a positioning table. Current dedicated breast PET systems have been shown to miss breast cancers near the chest wall. To overcome this limitation the detectors on the BRPET system have a unique slanted light guide designed to allow the system to obtain better images of the subject’s chest wall. The photon counting rate capability of the system was improved via an optimization process that tested multiple coincidence pair selections and timing window settings. This resulted in a peak noise equivalent count rate of 5.33 kcps, compared to 2.15 kcps prior to optimization. This was followed by a complete evaluation of imaging performance and the completion of a pilot human study testing the clinical viability of the system. To characterize the basic imaging performance of the BRPET system the measurements detailed by the National Electrical Manufacturers Association (NEMA) NU-4 2008 protocol were adapted. The scanner’s spatial resolution at the center of the FOV was measured to be 1.8, 1.7, and 1.9 mm FWHM in the axial, radial, and tangential directions, respectively. A total system sensitivity of 19.3% was observed. In addition, a set of unique tests was created to measure the system’s ability to reliably image close to the top of the examination table, i.e. to visualize posterior regions of the breast. The tests showed that the scanner ring can image up to a minimum distance of 6.25 mm from the top of the examination table under high contrast conditions. The pilot human study included 10 subjects who also underwent clinical contrast-enhanced MRI scans. There was a total of 11 biopsied lesions with 7 malignancies and 4 benign findings. The PET images were interpreted by two blinded NM radiologists while the clinical MRI images were interpreted by a blinded breast radiologist. The study results showed an average sensitivity and specificity of 92.5% and 100% respectively for BRPET, while the sensitivity and specificity for CE-MRI were 100% and 25% respectively. The pilot clinical evaluation of the system suggests it is a clinically viable system with the potential capability improve upon the specificity of current breast cancer imaging systems.
