About

Jae W. Song, MD, MS, is an Associate Professor of Radiology at the Hospital of the University of Pennsylvania and actively practices at Chester County Hospital, Penn Presbyterian Medical Center, Pennsylvania Hospital, and the Hospital of the University of Pennsylvania. His clinical expertise is in diagnostic neuroradiology, with a focus on interpreting brain, head and neck, and spine CT and MR imaging. He established and serves as the Director of the Penn Vessel Wall MR Imaging Program, integrating his clinical cerebrovascular expertise with the multidisciplinary efforts of the Penn Comprehensive Stroke Center. His research centers on understanding stroke mechanisms through establishing imaging biomarkers of atherosclerosis and inflammatory vasculitis using advanced non-invasive MR and CT imaging techniques. His work involves imaging intracranial and extracranial vasculature at submillimeter resolution to gain mechanistic insights into disease development and progression within vessel walls. His research aims to translate innovative diagnostic imaging technologies into personalized patient care, aligning with the goals of the Penn Stroke Center. Dr. Song holds a BA in Biology from the University of Pennsylvania, an MD from New York University School of Medicine, and an MS in Biostatistics and Clinical Epidemiology from the University of Michigan School of Public Health. He is actively involved in professional committees, serves as an Associate Editor for Radiology Advances, and has received numerous awards including a Fulbright Fellowship, NIH NRSA Award, and recognition for Research Excellence by the American Heart Association.

Research topics

• Medicine
• Gastroenterology
• Cardiology
• Internal medicine

Selected publications

• Explainable Machine‐Learning Model to Classify Culprit Calcified Carotid Plaque in Embolic Stroke of Undetermined Source

  Journal of Neuroimaging · 2026-01-01 · 1 citations

  articleOpen accessSenior author

  ABSTRACT Background and Purpose Embolic stroke of undetermined source (ESUS) may be associated with carotid artery plaques with <50% stenosis. Plaque vulnerability is multifactorial, possibly related to intraplaque hemorrhage (IPH), lipid‐rich necrotic core, perivascular adipose tissue (PVAT), and calcifications. Machine learning (ML)‐based plaque classification is increasingly popular but often limited in clinical interpretability by black‐box nature. We applied an explainable ML approach, using noncalcified plaque components and calcification features with the SHapley Additive exPlanations (SHAP) framework to classify plaques as culprit or nonculprit. Methods This was a retrospective, cross‐sectional study. Patients with unilateral anterior circulation ESUS with calcified carotid plaques in neck computed tomography (CT) angiography were analyzed. Calcification‐level features were derived from manual segmentations. Plaque‐level features were assessed by a neuroradiologist and by semi‐automated software. Plaques were classified as culprit if ipsilateral to stroke side. Eight classifiers were benchmarked, and a gradient‐boosted decision tree (CatBoost) was further tuned. SHAP explained model decisions. Results Seventy patients yielded 116 calcified plaques (270 calcifications). Model based on five plaque‐ and calcification‐level features achieved ROC‐AUC (receiver operating characteristic area under the curve) 0.79 and precision‐recall‐AUC 0.86, outperforming classification based on plaque thickness ≥3 mm (ROC‐AUC 0.59, p = 0.04) and IPH presence (ROC‐AUC 0.51, p = 0.003). SHAP identified plaque thickness and PVAT volume as the most influential features with potential thresholds of >2.6 mm and ≥112 mm 3 , respectively.f Conclusions ML model trained with noncalcified plaque and calcification features can classify culprit calcified carotid plaque better than conventional criteria. Using clinically interpretable features with SHAP, the model explained its decisions and suggested hypothesis‐generating thresholds.

  PublisherOA PDFDOI

• Neuroimaging features of cerebral air embolism: a matched case-control study

  American Journal of Neuroradiology · 2026-02-16

  articleOpen access

  BACKGROUND AND PURPOSE: Cerebral air embolism (CAE) is a rare but treatable cause of ischemic stroke. Clinically, CAE may be difficult to distinguish from stroke due to more typical thromboembolic causes, but accurate diagnosis is critical to initiate appropriate treatment. We aimed to define the imaging features of CAE by comparing MRI from patients with confirmed CAE to those in cardioembolic stroke due to atrial fibrillation (AF). MATERIALS AND METHODS: In a retrospective, matched case-control study, CAE cases from 2012-2023 were matched 1:2 by presenting NIHSS to control patients who had stroke due to AF and were not treated with thrombolytics or thrombectomy. MRIs were reviewed by a neuroradiologist blinded to group. The primary outcome was presence of pre-specified neuroimaging features on MRI. RESULTS: Fourteen patients with stroke due to CAE (median age 61, 64% female, median NIHSS 12) and 28 controls with stroke due to AF (median age 81, 43% female, median NIHSS 12) were included. The predominant infarction topography in CAE patients was gyriform in 86%, punctate in 7%, and wedge-shaped in 7%, whereas in patients with stroke due to AF the predominant infarction topography was wedge-shaped in 71%, punctate in 18%, and gyriform in 11% (p<0.001). CAE patients more often presented with multiple (93% versus 50%, p=0.007) and bilateral infarctions (79% versus 43%, p=0.05). Cortical borderzone involvement was more frequent in patients with CAE compared to those with AF (86% versus 25%, p<0.001). The presence of both predominantly gyriform infarction topography and cortical borderzone involvement had a 76.6% sensitivity and 96.4% specificity for CAE. CONCLUSIONS: CAE cause characteristic gyriform infarction patterns on MRI that are distinct from typical cardioembolic stroke. In addition, cortical borderzone predilection and multifocal infarctions were substantially more frequent in CAE. This constellation of findings, in the appropriate clinical context, should strongly suggest CAE as the mechanism of neurologic injury, and may facilitate timely identification of this uncommon but critical diagnosis.

  PublisherDOI

• Expert consensus on intracranial vessel wall MRI in cerebrovascular disease: Society for Magnetic Resonance Angiography recommendations

  European Radiology · 2026-02-13 · 2 citations

  article
  PublisherDOI

• RAPID CTA versus JLK LVO for large vessel occlusion detection: a pragmatic comparison of performance and common pitfalls

  Neuroradiology · 2026-03-10

  articleOpen access

  Rapid detection of large vessel occlusion (LVO) is crucial for improving outcomes of acute ischemic stroke. This study provides a real-world, head-to-head comparison of two commercial AI tools for automated LVO detection—RAPID CTA (vessel density-based) and JLK LVO (deep learning-based)—in a Korean stroke center. We retrospectively analyzed 176 consecutive patients with suspected stroke who underwent both CT angiography and CT perfusion. The performance of RAPID CTA and JLK LVO was compared against expert neuroradiologist consensus using the area under the receiver operating characteristic curve (AUROC). Misclassified cases (false positives [FPs] and false negatives [FNs]) were reviewed to determine their underlying causes. LVO was confirmed in 53 of 176 patients (30.1%). Both tools demonstrated high and comparable overall performance (AUROC 0.93 for both, p = 0.64). The causes for misclassifications were also consistent across both platforms. The most common cause of FPs was high-grade intracranial stenosis mimicking occlusion. The primary cause for FNs was the presence of well-developed collateral flow in distal occlusions, which masks the vessel cut-off. However, a matched-sensitivity analysis revealed different performance trade-offs; at a predefined threshold yielding 83% sensitivity, JLK LVO demonstrated higher specificity than RAPID CTA (0.96 vs. 0.89). Both RAPID CTA and JLK LVO are effective tools, but they exhibit distinct performance trade-offs. A clear understanding of each tool’s common pitfalls and performance trade-offs is crucial for clinicians to effectively integrate these AI results for optimal patient care.

  PublisherOA PDFDOI

• Prevalence of High-Risk CTA-Based Carotid Plaque-RADS Subtypes in Patients With Embolic Stroke of Undetermined Source

  Stroke · 2025-01-24 · 18 citations

  articleOpen access1st authorCorresponding

  BACKGROUND: A modified computed tomography angiography (CTA)–based Carotid Plaque Reporting and Data System (Plaque-RADS) classification was applied to a cohort of patients with embolic stroke of undetermined source to test whether high-risk Plaque-RADS subtypes are more prevalent on the ipsilateral side of stroke. With the widespread use of CTA for stroke evaluation, a CTA-based Plaque-RADS would be valuable for generalizability. METHODS: A retrospective observational cross-sectional study was conducted at a single integrated health system comprised of 3 hospitals with a comprehensive stroke center between October 1, 2015, and April 1, 2017. Patients with unilateral anterior circulation stroke and &lt;50% carotid stenosis on CTA were retrospectively identified. Maximum plaque thickness and ulceration were assessed by a neuroradiologist blinded to the stroke side. A semiautomated segmentation software measured intraplaque hemorrhage volumes. Modified CTA-based Plaque-RADS classification was defined as (1) no plaque, (2) plaque thickness &lt;3 mm, (3) plaque thickness ≥3 mm or ulcerated, and (4) plaque with intraplaque hemorrhage &gt;50 mm 3 irrespective of plaque thickness. High-risk plaque subtypes (Plaque-RADS 3 and 4) were compared with low-risk subtypes (Plaque-RADS 1 and 2). RESULTS: Ninety-four patients (55% women; median age, 66 years) were included. CTA-based Plaque-RADS categories for plaques ipsilateral to the stroke side were as follows: (1) 14.9%, (2) 42.6%, (3) 41.5%, and (4) 1.1%. Carotid plaques contralateral to stroke side were Plaque-RADS: (1) 21.3%, (2) 46.8%, (3) 31.9%, and (4) 0%. When compared with the contralateral side, plaques ipsilateral to the stroke side were significantly associated with high-risk Plaque-RADS subtypes in a mixed-effects logistic model adjusting for age and sex (adjusted odds ratio, 2.10 [95% CI, 1.20–3.71]; P =0.01). CONCLUSIONS: Carotid plaque ipsilateral to the stroke side was significantly associated with CTA-based high-risk Plaque-RADS subtypes in an embolic stroke of undetermined source cohort. A CTA-based Plaque-RADS classification may be useful for identifying potentially causative carotid plaque phenotypes in patients with embolic stroke of undetermined source.

  PublisherOA PDFDOI

• Segmentation of Leukoaraiosis on Noncontrast Head CT Using CT‐MRI Paired Data Without Human Annotation

  Brain and Behavior · 2025-06-01 · 1 citations

  articleOpen access

  OBJECTIVE: Evaluating leukoaraiosis (LA) on CT is challenging due to its low contrast and similarity to parenchymal gliosis. We developed and validated a deep learning algorithm for LA segmentation using CT-MRIFLAIR paired data from a multicenter Korean registry and tested it in a US dataset. METHODS: We constructed a large multicenter dataset of CT-FLAIR MRI pairs. Using validated software to segment white matter hyperintensity (WMH) on FLAIR, we generated pseudo-ground-truth LA labels on CT through deformable image registration. A 2D nnU-Net architecture was trained solely on CT images and registered masks. Performance was evaluated using the Dice similarity coefficient (DSC), concordance correlation coefficient (CCC), and Pearson correlation across internal, external, and US validation cohorts. Clinical associations of predicted LA volume with age, risk factors, and poststroke outcomes were also analyzed. RESULTS: The external test set yielded a DSC of 0.527, with high volume correlations against registered LA (r = 0.953) and WMH (r = 0.951). In the external testing and US datasets, predicted LA volumes correlated with Fazekas grade (r = 0.832-0.891) and the correlations were consistent across CT vendors and infarct volumes. In an independent clinical cohort (n = 867), LA volume was independently associated with age, vascular risk factors, and 3-month functional outcomes. INTERPRETATION: Our deep learning algorithm offers a reproducible method for LA segmentation on CT, bridging the gap between CT and MRI assessments in patients with ischemic stroke.

  PublisherOA PDFDOI

• Author response for "Segmentation of Leukoaraiosis on Noncontrast Head CT Using CT-MRI Paired Data Without Human Annotation"

  2025-05-10

  peer-review
  PublisherDOI

• Assessing Reliability, Reproducibility, and Adherence of Home Spirometry in Patients With Idiopathic Pulmonary Fibrosis: A Korean Multicenter Prospective Study

  American Journal of Respiratory and Critical Care Medicine · 2025-05-01

  articleSenior author

  Abstract Rationale: Idiopathic pulmonary fibrosis (IPF) is a progressive lung disease that requires regular lung function monitoring to assess disease progression. Forced vital capacity (FVC) is a main indicator of IPF progression, typically measured through on-site spirometry. However, frequent hospital visits may be burdensome for patients, and home spirometry can offer a potential alternative for continuous lung function monitoring. Although home spirometry has shown promise in chronic respiratory disease, data on its use among Korean patients with IPF are limited. We aimed to evaluate the reliability, reproducibility, and adherence of home spirometry compared to on-site measurements in a Korean IPF cohort. Methods: This multicenter prospective observational study enrolled 120 IPF patients from 12 medical institutes across South Korea. Participants performed home spirometry (SPROLENIS, JNBIO, South Korea) 10 times over five days at 12-hour intervals (8:00 AM and 8:00 PM) in each season for one year, following standardized training to ensure consistent device use. On-site spirometry was used as the reference. Reliability and agreement between home and on-site spirometry were assessed using intraclass correlation coefficients (ICCs) and Bland-Altman analysis. Reproducibility of home spirometry was evaluated using the standard deviation (SD) and coefficient of variation (CV, calculated as SD/mean × 100) for FVC values, while adherence was categorized as good (≥8 tests), fair (5-7 tests), or poor (≤4 tests). Results: Among all patients, the mean age was 68.4 years, with 85.0% being male and 77.5% having a history of smoking. Participants completed an average of 8.1 home spirometry tests each, with home FVC values showing a strong correlation with on-site FVC measurements (FVC % predicted: r = 0.834; FVC, L: r = 0.926; both p &amp;lt; 0.001) (Figure). Reliability analysis using the intraclass correlation coefficient (ICC) indicated excellent agreement between home and on-site FVC% predicted (ICC = 0.909; 95% CI: 0.889-0.937), supporting the robustness of home spirometry data. Bland-Altman analysis showed a mean difference of -8.2% in FVC % predicted, with 95% limits of agreement ranging from -22.9% to 6.5% predicted, suggesting satisfactory concordance. The mean SD for home FVC % predicted was 3.8%, with a CV of 6.4 ± 9.3%, demonstrating acceptable reproducibility. Adherence was high, with 70.8% of participants classified as good users (≥8 tests), 20.0% as fair users, and only 17.5% as poor users. Conclusions: Home spirometry could be a reliable and reproducible option for monitoring lung function in patients with IPF, potentially serving as an alternative to on-site measurements.

  PublisherDOI

• Association of Carotid Plaque Calcification Attenuation With Intraplaque Hemorrhage Volume: 3D‐Segmentation Analysis

  Journal of Neuroimaging · 2025-07-01

  articleOpen accessSenior authorCorresponding

  ABSTRACT Background and Purpose Despite the high prevalence of plaque calcifications in carotid atherosclerosis, the association between morphologic and attenuation features of calcifications and intraplaque hemorrhage (IPH) remains unclear. Methods Calcific carotid plaques were identified on neck computed tomography angiographies (CTAs) from patients meeting criteria for embolic stroke of undetermined source. Plaque calcifications were manually segmented using 3D‐Slicer to quantify features, including volume and attenuation (Hounsfield Units [HU]). IPH volume (IPHvol) was quantified using a semi‐automated software. A linear mixed regression model evaluated associations between calcification features and IPHvol, adjusting for sex, age, and cardiovascular risk factors. An interaction term between calcification volume and attenuation was included after dichotomizing attenuation (&gt;924HU) and volume (&gt;30 millimeter [mm] 3 ) as high versus low on the basis of median values. Results From 70 patients (median age 68 years, 50% female), 116 calcific plaques containing 269 plaque calcifications were analyzed. Adjusting for age, cardiovascular risk factors, and plaque calcification features, being female showed lower IPHvols compared to males (mean ratio 0.34, p = 0.002). A significant interaction between calcification volume and attenuation emerged ( p = 0.042). Among plaques with low plaque calcification volumes, plaques with low‐attenuation (&lt;924HU) calcifications showed 5.53 times higher IPHvols than plaques with high‐attenuation calcifications ( p = 0.003). Among plaques with high‐attenuation calcifications, plaques with high volumes of these calcifications showed 4.40 times higher IPHvols compared to low‐volumes of high‐attenuation calcifications ( p = 0.011). Conclusions Plaque calcification attenuation characteristics are associated with IPHvols. Understanding calcification patterns that correlate with IPH could enable clinicians to infer plaque instability from readily visible calcifications on CTA.

  PublisherOA PDFDOI

• Optimizing scan efficiency of T1-weighted imaging for whole-brain intracranial vessel wall imaging

  medRxiv · 2025-06-24

  preprintOpen accessCorresponding

  Background: Clinical intracranial vessel wall imaging (VWI) requires high spatial resolution leading to long scan times and artifacts. Purpose: To accelerate standard-of-care (SOC) 3D T1-weighted variable-flip-angle turbo-spin-echo (VFA-TSE) sequence with parallel imaging (Generalized Autocalibrating Partially Parallel Acquisitions, GRAPPA) using compressed sensing (CS) or Controlled Aliasing in Parallel Imaging Results in Higher Acceleration (CAIPIRINHA, CAIPI) with either standard or large field-of-view (FOV) configurations to reduce scan time, artifacts and accommodate head sizes. Study Type: Prospective study. Subjects: Ten healthy volunteers. Field Strength/Sequence: 3 Telsa, 20-channel head coil, T1-weighted VFA-TSE. Assessment: Accelerated sequences were compared to SOC GRAPPA (R=2), including standard FOV CAIPI (SFCAIPI, R=4), CS (SFCS7, R=7), and large FOV CS (LFCS7, R=7; LFCS10, R=10). Four neuroradiologists rated image quality (IQ) and signal-to-noise ratio (SNR) using a 4-point Likert scale. Scores of 3-4 were categorized as clinically interpretable. Lumen and wall diameters were measured. Statistical Analysis: Descriptive statistics are reported. McNemar's test compared proportions of IQ- and SNR-based clinically interpretable scans between relevant sequences of interest. Inter- and intra-rater reliabilities were calculated with Fleiss Kappa and weighted Cohen's Kappa, respectively. Lumen and wall diameters of the CS- and CAIPI-accelerated sequences were compared to SOC using paired t-tests. Results: SFCAIPI showed the lowest mean IQ and SNR scores. SFCS7 showed no significant difference in the proportion of IQ-based clinically interpretable scans compared to SFGRAPPA. When testing FOV, LFCS7 (35/40 scans; time of acquisition (TA)=3:45) showed a significantly higher proportion of IQ-based clinically interpretable scans compared to SFCS7 (27/40, p=0.03; TA=6:37). Upon increasing acceleration (R=10), there was no difference in the proportion of IQ-based clinically interpretable scans between LFCS7 and LFCS10 (36/40, p=0.65). Large FOV eliminated aliasing artifacts compared standard FOV (aliasing in 7 of 10 subjects). LFCS10 (TA=4:55) achieved a 50.6% reduction in TA relative to SFGRAPPA (TA=9:57). Conclusion: Large FOV CS VWI sequence with 10x acceleration achieved a 50.6% reduction in scan time while delivering image quality comparable to SOC standard FOV GRAPPA.

  PublisherOA PDFDOI

Recent grants

• NIH Grant F32AR058105

  NIH · $104k · 2012

Frequent coauthors

• Quy Cao

  University of Pennsylvania

  40 shared

• Kevin C. Chung

  Michigan Medicine

  36 shared

• Emi Takahashi

  Boston Children's Hospital

  28 shared

• Carly M. O’Donnell

  University of Pennsylvania

  27 shared

• Abigail R. Manning

  Penn Center for AIDS Research

  27 shared

• Melissa L. Martin

  University of Pennsylvania

  27 shared

• Daniel S. Reich

  National Institutes of Health

  27 shared

• Erin E. Longbrake

  Yale University

  26 shared

Awards & honors

• Fulbright Fellowship to Freiburg, Germany
• NIH NRSA Award
• Recognition for Research Excellence by the American Heart As…
• Andlinger Fellowship at the Medical University of Vienna, Au…