Predrag R. Bakic · University of Pennsylvania · PhdFit
Resume-aware faculty matching
Find professors who actually fit you
Upload your resume. PhdFit's six research agents compare your background with faculty profiles, recent publications, lab focus, and outreach opportunities, then rank professors with evidence you can review.
Resume-aware matchingEvidence from recent researchOutreach drafts included
Advisor Match Brief Resume uploaded
Resume signals5 found
NLPHCIFairnessGraph MLPyTorch
Six-agent analysisRunning
Axel · Candidate AnalystResume parsed
Nova · Professor ResearcherRecent papers
Echo · Match ExplainerEvidence linked
Atlas · Strategy AdvisorFit ranked
Lyra · Outreach WriterDraft queued
Juno · Meeting Prep CompilerNext step ready
SC
Dr. Sarah Chen
Stanford · NLP & interpretability
91
fit
Fairness + NLP projects overlap with her recent work.
Strong methods fit: model evaluation and interpretability.
Clear outreach angle for current trustworthy AI research.
Outreach angle
Ask how her lab is extending interpretability methods into fairness audits for real-world AI systems.
Predrag R. Bakic
· Ph.D.Verified
University of Pennsylvania · Rehabilitation Medicine
Predrag R. Bakic, Ph.D., is an Adjunct Associate Professor of Radiology at the University of Pennsylvania's Perelman School of Medicine. He is affiliated with the Department of Radiology at the Hospital of the University of Pennsylvania. Dr. Bakic holds a B.S. and M.S. in Electrical Engineering from the University of Belgrade, obtained in 1990 and 1995 respectively, and earned his Ph.D. in Electrical Engineering from Lehigh University in 2000. He further completed a postdoctoral fellowship in Medical Physics at Thomas Jefferson University in 2003. His research and professional focus are centered on medical physics and radiology, contributing to the academic and clinical missions of the university.
Breast density, defined as the proportion of breast tissue composed of dense fibroglandular tissue, is a crucial factor in assessing breast cancer risk and significantly impacts the visibility of lesions during mammography screening. We focus on distinguishing between high and low breast density by fine-tuning deep neural networks and on our joint deep and sparse approximation methodology. We evaluate the performance of these approaches in classifying high versus low breast density. Furthermore, we propose an example-based explainable AI approach denoted as Deep Sparse Reconstruct (DSR), which visualizes the most influential deep features in the corresponding mammogram region of interest, identified by non-negative sparse representations derived from a training dictionary. We also utilize established explainable AI techniques that visualize model predictions to facilitate comparisons. Our findings support that DSR enhances interpretability by extracting the key training features that contribute to predictions, it is compatible with various deep network architectures, and may contribute towards the development of trustworthy AI diagnostic workflows.
Purpose: Breast cancer is one of the leading causes of cancer-related deaths among women, and digital mammography plays a key role in screening and early detection. The radiation dose on mammographic exams directly influences image quality and radiologists' performance. We evaluate the impact of an image restoration pipeline-designed to simulate higher dose acquisitions-on the detectability of microcalcifications of various sizes in mammograms acquired at different radiation doses. Approach: The restoration pipeline denoises the image using a Poisson-Gaussian noise model, combining it with the noisy image to achieve a signal-to-noise ratio comparable with an acquisition at twice the original dose. We created a database of images using a physical breast phantom at doses ranging from 50% to 200% of the standard dose. Clustered microcalcifications were computationally inserted into the phantom images. The channelized Hotelling observer was employed in a four-alternative forced-choice to evaluate the detectability of microcalcifications across different sizes and exposure levels. Results: of the standard dose resulted in detectability levels comparable with those of images acquired at the standard dose. Moreover, images restored at the standard dose demonstrated detectability similar to those acquired at 160% of the nominal radiation dose, with no statistically significant differences. Conclusions: We demonstrate the potential of an image restoration pipeline to simulate higher quality mammography images. The results indicate that reducing noise through denoising and restoration impacts the detectability of microcalcifications. This method improves image quality without hardware modifications or additional radiation exposure.
Abstract Ultrasound Optical Tomography (UOT) combines the high-resolution imaging capability of ultrasound with measurements of light absorption and scattering properties of human tissue. In this work, UOT working at 794.2 nm wavelength and equipped with a spectral hole burning filter was used to image through 5 cm thick tissue phantoms with embedded cubic 12 × 12 × 12 mm 3 inclusions at a 2.5 cm depth. In contrast to earlier UOT works at tissue-relevant wavelengths, these inclusions have orders of magnitude lower absorption and actually mimic the optical properties of human breast tissue with various lesions. Phantoms with inclusions of increasing reduced scattering (inclusions 9.37–14.1 cm −1 , background 7.83 and 8.5 cm -1 ) and absorption (inclusions 0.061–0.086 cm −1 , background 0.044 and 0.045 cm -1 ) coefficients were investigated. In the UOT images, the contrast-to-noise ratios varied between 3.87 ± 1.71 and 6.25 ± 2.96, and the inclusions could be easily identified by eye. This indicates that the UOT technique has the potential for spatially resolved imaging and optical data acquisition through at least 5 cm of soft tissue. Our findings suggest that UOT equipped with a spectral hole burning filter is a promising technique for breast tumour imaging as well as qualitative and quantitative characterization of their optical properties.
Rapid adoption of artificial intelligence methods in breast imaging research emphasizes the need for large, appropriately curated image databases for development and validation. For digital breast tomosynthesis (DBT), there are few public databases with only limited lesion annotation. Recently, we have developed Malmö Breast ImaginG (M-BIG), a large database of 104 791 women screened at Skåne University Hospital, Malmö. M-BIG also includes all images from the Malmö Breast Tomosynthesis Screening Trial, MBTST of 14 848 women, with 139 biopsy-confirmed cancers from DBT screening. To annotate lesions in M-BIG, we designed a semi-automated custom software tool for DBT, and corresponding digital mammography (DM) images. A reader manually draws an outline; or marks nodes around the lesion which are automatically connected by an edge-following algorithm. Our custom tool enables detailed annotation of DBT and DM lesions, as opposed to the rectangular regions present in other published material, allowing extensive evaluation of tumor segmentation, and analysis of size and shape descriptors.
We propose analytical expressions for ultrasound optical tomography and verify them in Monte Carlo simulations. The proposed hybrid analytical-statistical Monte Carlo model enables quantitative reconstruction of up to 5-centimetre deep tissue absorption inhomogeneities.
Abstract Simulation methods in breast imaging offer advantages over clinical trials in terms of improved reproducibility, reduced need for patient exposure to radiation, increased flexibility, and more clearly defined ground truth. Simulation also allows for improved representation of anatomical variations and variations in acquisition parameters and breast positioning related to multimodality imaging. The increasing use of virtual clinical trials (VCTs) to assess breast imaging systems has introduced a demand to optimize protocols for simulation studies. This work will contribute to developing standards for evaluation tools for 3D/4D breast imaging systems and will ultimately reduce the reliance on clinical trials for emerging systems. This report reviews key aspects of VCTs, including the simulation of realistic breast anatomy, the generation of synthetic images from virtual phantoms, the use of model observers to assess imaging system performance, and methods to analyze observer outputs. Each section reviews the state of the science and recommends approaches for accomplishing tasks related to the individual aspects of VCTs. The report also reviews the experience of designing and using a simulation approach from the industrial and regulatory perspective. Finally, future steps in the development of VCTs are suggested. breast cancer imaging, evaluation of imaging systems, virtual trials
In this study, NaCl pellets read with optically stimulated luminescence were evaluated for their potential use as a point dosemeter in breast x-ray imaging. Dosimetry with NaCl pellets has previously been applied to various environmental and medical settings. NaCl pellets have potential in clinical breast dosimetry because they can enable multiple point measurements to be simultaneously conducted in a cost- and time-efficient manner. Using two digital mammography imaging systems, the air kerma response of the NaCl pellets for a standard breast setup was investigated. The air kerma response was observed to be linear, and mathematical fits were successfully used to estimate the cumulative incident air kerma during digital breast tomosynthesis. Deviations from a reference digital dosemeter were 6% and 8% for the two mammography systems, respectively. Measurements conducted at different angles of exposure showed that the NaCl pellets had no angular dependency in the range ± 15°. Finally, the uniformity of the beams was confirmed to avoid possible errors due to the uncertainty of the dosemeters' positions in the exposure field.
BACKGROUND: Virtual imaging trials (VITs) are in silico studies that simulate medical imaging and disease processes, offering a cost-effective and reproducible addition to traditional imaging trials. While VITs are well established in breast imaging, most existing implementations simulate imaging based on static anatomical models, capturing only a single time point. This limits their ability to study time-dependent processes such as tumor progression or breast density and composition changes over time. PURPOSE: We introduce STELLA-R (Simulation of Temporal Evolution and LongitudinaL studies of breast Anatomy in Radiology), the first framework aimed at performing longitudinal virtual imaging trials in breast imaging. STELLA-R is designed to simulate temporal changes in breast anatomy, density, and lesion development across a virtual population of women. METHODS: Our simulation pipeline consists of five modules. The population creator module generates realistic virtual cohorts based on real-world data from approximately 25 000 women, modeling multivariate distributions of age, breast shape, and breast density. The phantom creator and lesion creator modules enable detailed specification of breast and lesion characteristics, utilizing Perlin noise computational algorithms to replicate tissue appearance. The tumor location is assigned in the lesion insertion module. To simulate temporal changes in the breast, we used real-world data from two consecutive screening rounds. This enabled realistic modeling of mammographic density evolution, breast volume changes, and tumor growth. Different breast densities were achieved by adjusting threshold values applied to the Perlin noise, which determines the amount of tissue structure. Temporal changes of parenchyma were simulated by gradually varying the threshold values. Tumor progression was simulated by increasing lesion size according to growth rates sampled from real-world data. Lastly, the Image Generation module integrates multiple external software components for mammographic image formation, including noise and scatter simulation and image reconstruction. In this study, we simulated digital breast tomosynthesis (DBT) images of our phantoms using open-source tools. Our simulation framework is modular and can be extended to support additional imaging modalities. RESULTS: We demonstrate case examples of virtual women at ages 40, 57, and 74, reflecting Swedish screening intervals, and report simulated changes in volumetric breast density over time (14%, 9%, and 6%, respectively). The breast density is modeled with a mean accuracy of < 2% compared to target values. Additionally, we illustrate lesion progression across multiple time points, assuming a tumor doubling time of 282 days. Our fitted models accurately capture correlations between age, breast volume, density, and annual changes. CONCLUSION: STELLA-R pipeline provides a novel foundation for evaluating long-term screening strategies, imaging, and risk models in a controlled and customizable manner using longitudinal VITs.
