MSBI Thesis & Abstracts 2024-2025

Rocelle Evangelista

Thesis Title: Diffusion MRI Biomarkers of RNS Efficacy in Epilepsy

Advisor: Ankit Khambhati, PhD

Abstract: Rationale: Responsive neurostimulation (RNS) is a promising therapy for medically refractory epilepsy (MRE), yet clinical outcomes remain highly variable and difficult to predict. Identifying preoperative biomarkers that forecast treatment response could improve patient selection and inform neuromodulation strategies. We investigated whether structural and diffusion MRI features from limbic white matter pathways and mesial temporal structures could predict seizure reduction with hippocampal-targeted RNS.Methods: We retrospectively analyzed 41 patients who underwent hippocampal-targeted RNS implantation at UCSF. Diffusion tensor imaging (DTI) metrics and volumetric measures from mesial temporal and limbic regions were analyzed. Clinical response was quantified as time weighted percent seizure reduction across follow up visits. ElasticNet regression was employed to identify imaging features associated with response, with performance assessed using leave-one-out cross-validation and bootstrapped confidence intervals.Results: Larger right amygdala volume (β = 5.75, p = 0.020) and high right uncinate fasciculus FA (β = 4.79, p = 0.015) were positively associated with seizure reduction, while lower FA in the left (β = –4.47, p = 0.008) and right (β = –3.48, p = 0.032) fornix was negatively associated. The full training model explained 46.5% of the variance, but cross-validation revealed limited generalizability (CV R² = –0.83 ± 2.57).Conclusions: Preoperative morphometric and white matter microstructural features show biologically plausible associations with RNS therapy response, particularly within mesial temporal and frontolimbic networks. Though not individually predictive, these findings suggest imaging biomarkers could inform neuromodulation strategies.

Nicholas Finke

Thesis Title: A Large Language Model for Liver Lesion Characteristic Extraction and Longitudinal Tracking from LI-RADS Reports

Advisor: Yang Yang, PhD

Abstract: Background: Large language models (LLMs) can be utilized to summarize radiologist-written reports for abdominal MRI and CT scans of liver lesions suspicious for hepatocellular carcinoma (HCC), and identify and track the growth of lesions over time. The standardized liver lesion reporting method, Liver Imaging Reporting and Data Systems (LI-RADS), allows for consistent representation of lesion characteristics which aids in the consistency of feature extraction.Objective: This study evaluates the ability of an LLM to compile qualitative radiologist reports into a structured database containing liver lesion identification and developmental features, and subsequently using those features to identify and track specific lesions across scans. Methods: This retrospective study included patients at risk for HCC who underwent abdominal CT or MRI scans that contained a LI-RADS 5 lesion. The radiologist reports of 108 scans from 15 patients were inputted to an LLM (GPT-4.1) alongside a standardized CT/MRI LI-RADS diagnostic algorithm. The LLM first recorded the relevant LI-RADS features from each report, then assigned a unique lesion ID for each lesion that developed into LI-RADS 5. The result was compared to human-based identification and tracking. Results: The LLM utilizing LI-RADS classification standards had an accuracy of 99.4% in identifying LI-RADS traits and assigned score. LI-RADS 5 longitudinal lesion tracking accuracy across scans was lower at 42.6%.Conclusion: The LLM is effective in automated feature extraction and lesion classification, but even with the standardized LI-RADS reporting system, inconsistent lesion descriptions across scans and limited standardized lesion tracking documentation limits the effectiveness of longitudinal tracking.

Bhavna Gundamaraju

Thesis Title: Resting state functional magnetic resonance imaging in presymptomatic and symptomatic genetic prion disease

Advisor: Michael Geschwind, MD, PhD

Abstract: Genetic prion diseases (gPrDs) are rare, fatal neurodegenerative disorders caused by pathogenic variants in the PRNP gene. Identifying early markers of network dysfunction is critical for advancing early detection and intervention strategies. This study investigates default mode network (DMN) functional connectivity alterations in presymptomatic gPrD carriers, with a focus on differences between faster- and slower-progressing subgroups. Resting-state fMRI data was acquired from presymptomatic carriers, symptomatic participants, and matched healthy controls. Seed-to-voxel analyses were performed to compare DMN connectivity across groups, with statistical significance determined after correction for multiple comparisons. Presymptomatic carriers exhibited distinct connectivity alterations based on progression rate. Faster-progressing individuals showed reduced posterior DMN connectivity, centered in the posterior cingulate and precuneus, while slower-progressing carriers displayed increased anterior DMN connectivity within the medial prefrontal cortex. Symptomatic participants demonstrated widespread DMN disruption relative to controls, consistent with progressive network breakdown. Alterations in DMN connectivity are detectable before symptom onset and differ according to progression rate, supporting their potential as early biomarkers in gPrD. While constrained by sample size and limited longitudinal follow-up, these findings highlight the value of resting-state fMRI in presymptomatic stages and underscore the need for larger, multicenter, longitudinal studies to clarify the trajectory of functional network changes in gPrD.

Bhakti Kulkarni

Thesis Title: PET/MRI Assessment of Patellofemoral Osteoarthritis Pain Induced by Exercise

Advisor: Richard Souza, PhD, PT

Abstract: Background: Patellofemoral Joint Osteoarthritis (PFJ OA) is a chronic articular disease in the knee joint characterized by the degradation of cartilage and pain. Previous studies have found degraded cartilage in asymptomatic individuals while others suggest a correlation between inflammation and pain. 18F-fluorodeoxyglucose ([18F]-FDG), a glucose analog radiotracer, reflects glucose hypermetabolism and may suggest an influx of proinflammatory cells due to the need of oxygen consumption during the inflammatory process. This study aims to highlight the anatomical structures in the PFJ that experience hypermetabolism which may indicate the presence of inflammation before and after exercise using positron emission tomography (PET) with [ 18F]-FDG and magnetic resonance imaging (MRI). Methods: Five subjects were enrolled: one control subject and four subjects diagnosed with PFJ OA. Images were acquired before and after subjects performed 25 single-legged squats. Eight knee regions were segmented: infrapatellar fat pad, patellar tendon, meniscus, patella, tibia, femur, lateral PFJ capsule and synovium, and medial PFJ capsule and synovium. Spherical regions of interest (ROIs) were defined in each region to calculate local standard uptake values (SUVpeak, and SUVmax). Results: [ 18F]-FDG uptake was present in all tissues. Post-exercise scans showed increased [ 18F]-FDG uptake across most PFJ regions in most subjects. In all subjects on average, had a pre-squat SUVmax value of 0.859 and post-squat exercise SUVmax value of 0.959 in most regions. Then, all subject on average had a pre-squat SUVpeak value of 0.850 and a post-squat SUVpeak value of 0.918 in most regions. Control SUV values consistently remained lower than PFJ OA subjects. Conclusion: From this preliminary data, considerable variability was noted in uptake values across knee joint tissues in both patients and controls. In general, greater tracer uptake was noted in all tissues in subjects with PFJ OA when compared to the control subject, indicating greater glucose metabolism and suggesting hypermetabolic inflammation. Post-exercise, the tibia and medial synovium demonstrated the highest change in SUVmax and SUVpeak, possibly indicating the presence of exercise-induced inflammation.

Alexandra Wu

Thesis Title: MRI Changes in Presymptomatic Genetic Prion Disease Patients

Advisor: Michael Geschwind, MD, PhD

Abstract: Purpose: Early detection of symptomatic prion disease (PrD) is critical because it allows for earlier intervention, management, and treatment of symptoms. The goal of this project is to evaluate the utility of imaging biomarkers in identifying microstructural abnormalities within genetic prion disease (gPrD) patients prior to symptom onset. This was accomplished by comparing cross-sectional and longitudinal MRI (standard T1-weighted (T1w) and diffusion tensor imaging (DTI)) of healthy controls (HC) and presymptomatic (presymptomatic) gPrD carriers to assess for group-wise and individual-level abnormalities; and further analysis evaluated correlations between imaging abnormalities and clinical abnormalities of disease progression to investigate imaging’s potential as an early biomarker. Materials and Methods: This project utilizes standard structural T1w MRI and DTI images of HCs (recruited by the UCSF Memory and Aging Center) and presymptomatic gPrD participants (recruited by the Geschwind Lab) that have been continuously acquired since 2005. The Siemens MAGNETOM Trio 3.0T MRI scanner was used before approximately February 2016 and the Siemens MAGNETOM Prisma Fit 3.0T MRI scanner has been used since approximately October 2016. MATLAB’s Statistical Parametric Mapping (SPM) software and MRTrix3 software was used to process these images. Results: Cortical and subcortical MD was significantly decreased in presymptomatic Fast gPrD and the E200K subcohort compared to HC at both first and last presymptomatic MRI, particularly in temporal regions. At first presymptomatic Fast gPrD MRI, decreased MD was seemingly lateralized to the left hemisphere; but at the final presymptomatic Fast gPrD MRI, more regions showed left and right hemisphere involvement, strengthening bilateral significance. For example, at first presymptomatic Fast gPrD MRI the middle temporal gyrus p = 0.048 and at last presymptomatic Fast gPrD MRI the middle temporal gyrus p = 0.0087. There were also significant positive correlations between GM volume Z-scores (from both first and last presymptomatic Fast gPrD MRI) and executive-function neurocognitive composite scores for both the fast gPrD and the E200K groups. Conclusion: DTI shows promise as a potential imaging biomarker, having shown significant differences in gray matter mean diffusivity in presymptomatic Fast gPrD subjects compared to controls. This supports that microstructural brain changes occur before symptom onset (and can be detected by DTI) and may possibly coincide with clinical changes.

Cheng Xue

Thesis Title: Identification of Novel CD46-Binding Peptides for PET Imaging of Prostate Cancer Using Phage Display

Advisor: Robert Flavell, MD, PhD

Abstract: Prostate-specific membrane antigen (PSMA) PET is central to prostate cancer imaging, yet it shows reduced sensitivity in PSMA-low disease, including neuroendocrine variants. To develop CD46-targeted peptide tracers suited for faster imaging, we carried out four rounds of phage display with a Ph.D.-12 library against recombinant CD46. Rounds alternated between bead-based (1, 3) and plate-based (2, 4) formats, with antigen reduced from 25 to 10 µg and wash stringency increased (Tween-20 from 0.1% to 0.5%). Phage enrichment was evident by titer: amplified outputs rose from 1.8×10¹⁰ pfu/mL in round 1 to 9×10¹¹ pfu/mL in round 3, then measured 6×10⁹ pfu/mL in round 4. Next-generation sequencing showed contraction of library diversity from 6,623 unique peptides (50,804 total reads) in round 1 to 427 (3,873 reads) in round 4. Normalized-frequency trends highlighted strong enrichment of Peptide A (2.08%→17.84, rounds 1→4) and a late rise of Peptide B (0.44% in round 3 to 6.82% in round 4), while Peptide C declined. PHASTpep ranking with MEME/STREME analysis identified early enrichment of Motifs A–C, followed in later rounds by a cysteine-enriched signature (Motif F) and a recurrent core (Motif H). AlphaFold models suggested a plausible CD46-binding pose for peptide, whereas control peptide did not engage the surface. Ten peptides were synthesized for surface plasmon resonance validation, providing a short list of CD46-binding candidates for tracer development in PSMA-low prostate cancer.

Jiuyi Zhang

Thesis Title: Validation and Optimization of a Quantitative Susceptibility Mapping (QSM) Pipeline for Paramagnetic Rim Lesions in Multiple Sclerosis

Advisor: Roland Henry, PhD

Abstract: Quantitative Susceptibility Mapping (QSM) provides valuable sensitivity to paramagnetic rim lesions (PRLs) in multiple sclerosis, but the influence of algorithmic choices within the processing pipeline remains unclear. This study systematically compared two phase unwrapping methods (ROMEO and PRELUDE), two background field removal strategies (V-SHARP and RESHARP), and dipole inversion with iLSQR to assess their impact on lesion visibility and contrast stability.ROMEO achieved rapid phase unwrapping (20–40 seconds per dataset) compared to the several hours required by PRELUDE, yet PRELUDE produced clearer boundaries for small lesions, particularly in anatomically complex regions. For background field removal, V-SHARP provided greater global stability across brain regions, while RESHARP enhanced local lesion-to-background discrimination, especially within the corpus callosum and basal ganglia. Case-specific analyses further revealed that lesion distribution strongly influenced outcomes: when lesions were widely distributed across white matter, pipeline effects were minimal, whereas lesions clustered near iron-rich structures such as the basal ganglia amplified pipeline-dependent differences, yielding significant QSM-level effects.In summary, ROMEO offers major efficiency advantages, but PRELUDE remains essential for small-lesion analysis. Likewise, V-SHARP ensures stable global contrast, while RESHARP is more effective in certain anatomical contexts. These findings underscore the need for pipeline selection tailored to lesion characteristics and anatomical location.