Atrial fibrillation (AFib) is the most common sustained cardiac arrhythmia, contributing to significant morbidity and mortality worldwide. Patient-specific computational models of the left atrium are currently studied to predict characteristics of reentrant activity that promotes fibrillation. However, current models’ patient-specificity is limited to anatomical structure and the distribution of disease-related remodeling (fibrosis), whereas electrical properties of cells and tissue are based on literature values. In cases where patients are clinically known to present with either AFib or atrial flutter (AFl), this lack of personalization can lead to inaccuracies in simulation outcomes (e.g., AFib-like behavior in simulations for a patient who actually had AFl, or vice-versa). My goal was to derive parameter sets that favor the initiation of one type of arrhythmia or the other (AFib or AFl). Ten fibrotic left atria were reconstructed from late-gadolinium enhanced (LGE)-MRI scans and the bioelectric parameter space (comprising ion channel expression levels and impulse propagation rates) was explored using a Taguchi L27 Design of Experiments (DoE) approach. Arrhythmias were induced by initializing four atrial regions to different phases of the action potential under each parameter permutation. I ran 300 simulations and manually classified each arrhythmia episode as either AFib- or AFl-like based on prior definitions. I pinpointed a pro-AFl parameter set – bioelectrical conditions under which 89% of all induced arrhythmias were AFl and only 11% were AFib. The pro-AFib parameter set in these preliminary simulations was comparatively less robust (61% vs. 39% for AFib vs. AFl inductions, respectively). My future work on this project will establish stronger relationships between model configurations and simulation outcomes by probing a wider array of possible parameters in a larger population of patient-specific models. Data from the present study will guide future simulations to accurately tailor models to represent the arrhythmic state in patients predisposed to AFl.

Evaluating the risk of developing breast cancer is an important aspect of cancer care as it can allow for more tailored screening strategies and preventative therapies. Clinicians use multiple measures to determine a patient’s risk of developing breast cancer, including breast density on mammography and genetic mutations. Background parenchymal enhancement (BPE) on magnetic resonance imaging (MRI) has shown promise to improve stratification of breast cancer risk in women at high-risk of cancer development. BPE is the increase in signal intensity of normal breast tissue on dynamic contrast-enhanced (DCE) MRI after the administration of contrast agent. Despite BPE having an association with an increased risk of breast cancer development, the biological basis of this increased enhancement is unknown. The aim of this study is to investigate what biologically drives BPE by connecting quantitative MRI measurements with pathological markers from normal breast tissue. Our study cohort includes women that received prophylactic mastectomies and DCE-MRI scans acquired ≤1 year before surgery. From mastectomy specimens, pathological measures of COX-2, VEGF, and Ki-67 are used to measure inflammation, vascular recruitment, and proliferation, respectively. To quantify BPE, I used in-house software to correct pre-contrast images using N4 bias field correction and segment the whole breast. I then applied the breast mask to the pre-contrast MRI and used fuzzy c-means clustering to automatically segment fibroglandular tissue (FGT) from surrounding fat, generating an FGT mask. This mask was then applied to the DCE-MRI series, which includes pre- and post-contrast images, to calculate BPE, which is the mean percent enhancement across FGT. As part of ongoing work, I will obtain more specific measurements in quadrants of the breast from which the pathology specimen was derived. I will then correlate BPE measurements to the pathology measures to determine if any associations exist between BPE and inflammation, vascular recruitment, and proliferation.

Diffusion-weighted imaging (DWI) shows great potential for improving breast cancer detection and diagnosis. Primary findings from the ECOG-ACRIN A6702 multi-site, multi-vendor clinical trial indicate that DWI apparent diffusion coefficient (ADC) values may help reduce false positives and unnecessary biopsies. Gradient nonlinearity (GNL) correction was previously found to improve the accuracy of ADC mapping within and across MRI vendor systems. In this study, we evaluated the impact of GNL correction on breast lesion ADC measures in the A6702 dataset. The dataset comprised 81 suspicious breast lesions (28/81 malignant) in 67 women. Standardized DWI scans were acquired across 9 different MRI scanners. ADC maps were created from DWI scans, and ADC values were measured for each lesion. Direction-averaged GNL correction maps were constructed based on scanner-specific gradient specifications. ADC map correction was then performed through pixel-wise scaling by the GNL correction maps using custom software developed in MATLAB. Lesion ADCs before and after GNL correction were compared using a two-tailed z-test. ADC diagnostic performance (benign vs. malignant) was evaluated using area under the receiver-operating-characteristic-curve (AUC), and optimal ADC cutoffs were chosen to maximize specificity while maintaining 100% sensitivity. GNL-corrected lesion ADCs were significantly lower than uncorrected ADCs (1.12±0.29 vs 1.17±0.30x10-3mm2/s, p<0.001). GNL error in lesion ADCs varied across gradient systems (mean ∆ADCvendorA=0.14±0.08, ∆ADCvendorB=0.03±0.02, ∆ADCvendorC =0.004±0.01, p<0.001). GNL correction produced a slightly lower optimal ADC cutoff (1.33 vs. 1.35x10-3mm2/sec). However, no overall difference in diagnostic performance was detected: AUCuncorrected=0.78 (95% CI 0.68-0.88), AUCcorrected=0.79 (95% CI:0.69-0.89), p=0.22, and 18% potential biopsy reduction for both. This study showed GNL substantially affects lesion ADC measures, with significant variability across different vendor platforms. These findings suggest that GNL correction should be implemented to ensure uniformity and consistency in diagnostic breast lesion ADC measures across MRI platforms, especially for multi-center clinical studies.