Shawn Lyo

Shawn Lyo

Auto-generated profile, based on publicly available information

Shawn Lyo

Department of Radiology, Hospital of the University of Pennsylvania, Philadelphia, USA

Research Focus & Metrics

Academic Career Length

7Years (2020-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
19 August 2024 in Journal of Imaging Informatics in Medicine
Verified
Journal Article
Open access
PDF file
SLShawn Lyo
Shawn Lyo
SMSuyash Mohan
Suyash Mohan
AHAlvand Hassankhani
Alvand Hassankhani
ANAbass Noor
Abass Noor
FDFarouk Dako
Farouk Dako
TCTessa Cook
Tessa Cook
Expert feedback on trainees’ preliminary reports is crucial for radiologic training, but real-time feedback can be challenging due to non-contemporaneous, remote reading and increasing imaging volumes. Trainee report revisions contain valuable educational feedback, but synthesizing data from raw revisions is challenging. Generative AI models can potentially analyze these revisions and provide structured, actionable feedback. This study used the OpenAI GPT-4 Turbo API to analyze paired synthesized and open-source analogs of preliminary and finalized reports, identify discrepancies, categorize their severity and type, and suggest review topics. Expert radiologists reviewed the output by grading discrepancies, evaluating the severity and category accuracy, and suggested review topic relevance. The reproducibility of discrepancy detection and maximal discrepancy severity was also examined. The model exhibited high sensitivity, detecting significantly more discrepancies than radiologists (W = 19.0, p < 0.001) with a strong positive correlation (r = 0.778, p < 0.001). Interrater reliability for severity and type were fair (Fleiss’ kappa = 0.346 and 0.340, respectively; weighted kappa = 0.622 for severity). The LLM achieved a weighted F1 score of 0.66 for severity and 0.64 for type. Generated teaching points were considered relevant in ~ 85% of cases, and relevance correlated with the maximal discrepancy severity (Spearman ρ = 0.76, p < 0.001). The reproducibility was moderate to good (ICC (2,1) = 0.690) for the number of discrepancies and substantial for maximal discrepancy severity (Fleiss’ kappa = 0.718; weighted kappa = 0.94). Generative AI models can effectively identify discrepancies in report revisions and generate relevant educational feedback, offering promise for enhancing radiology training.

1 July 2025 in Journal of the American College of Radiology
Verified
Journal Article
Open access
STSatvik Tripathi
Satvik Tripathi
DADana Alkhulaifat
Dana Alkhulaifat
SLShawn Lyo
Shawn Lyo
RSRithvik Sukumaran
Rithvik Sukumaran
BLBolin Li
Bolin Li
TCTessa S. Cook
Tessa S. Cook
Large language models (LLMs) are reshaping radiology through their advanced capabilities in tasks such as medical report generation and clinical decision support. However, their effectiveness is heavily influenced by prompt engineering-the design of input prompts that guide the model's responses. This review aims to illustrate how different prompt engineering techniques, including zero-shot, one-shot, few-shot, chain of thought, and tree of thought, affect LLM performance in a radiology context. In addition, we explore the impact of prompt complexity and temperature settings on the relevance and accuracy of model outputs. This article highlights the importance of precise and iterative prompt design to enhance LLM reliability in radiology, emphasizing the need for methodological rigor and transparency to drive progress and ensure ethical use in health care.

15 July 2025 in Journal of the American College of Radiology
Verified
Journal Article
Open access
SLShawn K. Lyo
Shawn K. Lyo
TCTessa S. Cook
Tessa S. Cook
Radiology education is challenged by increasing clinical workloads, limiting trainee supervision time and hindering real-time feedback. Large language models (LLMs) can enhance radiology education by providing real-time guidance, feedback, and educational resources while supporting efficient clinical workflows. We present an interpretation-centric framework for integrating LLMs into radiology education subdivided into distinct phases spanning predictation preparation, active dictation support, and postdictation analysis. In the predictation phase, LLMs can analyze clinical data and provide context-aware summaries of each case, suggest relevant educational resources, and triage cases based on their educational value. In the active dictation phase, LLMs can provide real-time educational support through processes such as differential diagnosis support, completeness guidance, classification schema assistance, structured follow-up guidance, and embedded educational resources. In the postdictation phase, LLMs can be used to analyze discrepancies between trainee and attending reports, identify areas for improvement, provide targeted educational recommendations, track trainee performance over time, and analyze the radiologic entities that trainees encounter. This framework offers a comprehensive approach to integrating LLMs into radiology education, with the potential to enhance trainee learning while preserving clinical efficiency.

3 May 2020 in European Heart Journal - Case Reports
Verified
Journal Article
Open access
PDF file
SLShawn Lyo
Shawn Lyo
JMJeremy Miles
Jeremy Miles
JMJay Meisner
Jay Meisner
MGMark Guelfguat
Mark Guelfguat
Background: Glycogen storage disease type IV (GSD IV; Andersen’s disease) is a rare autosomal recessive disease caused by mutation in the GBE1 gene. Presentation of GSD IV varies on a continuum of severity and symptomatology ranging from neonatal death to mild adult-onset disease with variable involvement of hepatic, muscular, neurologic, dermatologic, and cardiac systems. Cardiomyopathy seen in GSD IV is also heterogeneous and its appearance on cardiac magnetic resonance imaging (CMR) is rarely described. Case summary: A 29-year-old man without previous medical history was admitted to our facility multiple times over 2 years for focal sensorimotor deficits, gout arthropathy, chronic hyperlactataemia and hyperuricaemia, and severe decompensated non-ischaemic cardiomyopathy complicated by episodes of thromboembolic organ infarction. Echocardiography and CMR showed severe biventricular failure with the presence of intraventricular thrombi with increased right ventricular trabeculation and absent late gadolinium enhancement. He underwent muscle biopsy which showed prominent glycogen in skeletal muscle followed by genetic testing showing a single heterozygous splicing mutation c.993-1G>T found at the junction of intron 7 and exon 8 of the GBE1 gene which had not previously been reported and was predicted to be pathologic. He was referred to a tertiary care centre with glycogen storage disease specialists but expired prior to establishing care at that facility. Discussion: Discovery of GSD IV in our patient was unexpected due to a highly variant clinical presentation. Our case stresses the clinical heterogeneity of GSD IV and the importance of genetic sequencing studies in the evaluation of potential glycogen storage disease.

1 July 2022 in British Journal of Radiology
Verified
Journal Article
Closed access
SLShawn Lyo
Shawn Lyo
LTLuis Octavio Tierradentro-García
Luis Octavio Tierradentro-García
AVAngela Nicole Viaene
Angela Nicole Viaene
MHMisun Hwang
Misun Hwang
Objective: To assess the feasibility of visualizing lenticulostriate vessels (LV) using a linear high-resolution ultrasound probe and characterize LV morphology to determine whether morphological alterations in LV are present in neonatal hypoxic–ischemic encephalopathy (HIE) as compared to the unaffected infants. Methods: We characterized LV by their echogenicity, width, length, tortuosity, and numbers of visualized stems/branches in neurosonographic examinations of 80 neonates. Our population included 45 unaffected (non-HIE) and 35 with clinical and/or imaging diagnosis of HIE. Of the neonates with clinical diagnosis of HIE, 16 had positive MRI findings for HIE (HIE+MRI) and 19 had negative MRI findings (HIE-MRI). Annotations were performed twice with shuffled data sets at a 1-month interval and intrarater reliability was assessed. Focused comparison was conducted between non-HIE, HIE+MRI and HIE-MRI neonates whose images were acquired with a high frequency linear transducer. Results: Studies acquired with the two most frequently utilized transducers significantly differed in number of branches (p = 0.002), vessel thickness (p = 0.007) and echogenicity (p = 0.009). Studies acquired with the two transducers also significantly differed in acquisition frequency (p < 0.001), thermal indices (p < 0.001) and use of harmonic imaging (p < 0.001). Groupwise comparison of vessels imaged with the most frequently utilized transducer found significantly fewer branches in HIE + MRI compared to HIE-MRI negative and non-HIE patients (p = 0.005). Conclusion: LV can be visualized in the absence of pathology using modern high-resolution neurosonography. Visualization of LV branches varies between HIE + MRI, HIE-MRI neonates and controls. Advances in knowledge: High-resolution neurosonography is a feasible technique to assess LV morphology in healthy neonates and neonates with HIE.

23 August 2025 in Radiology Advances
Verified
Journal Article
Open access
SLShawn K Lyo
Shawn K Lyo
SMSuyash Mohan
Suyash Mohan
MHMichael J Hoch
Michael J Hoch
VPVivek P Patel
Vivek P Patel
RKRobert M Kurtz
Robert M Kurtz
AHAlvand Hassankhani
Alvand Hassankhani
Background: Magnetic resonance imaging (MRI) is a cornerstone of neuroimaging but is limited by lengthy acquisition times, which can lead to motion artifacts, patient discomfort, and delayed care. Deep learning reconstruction is an emerging technology that can offer image acquisition acceleration while maintaining image quality. Purpose: To compare image quality and acquisition efficiency between deep learning–accelerated vs conventional MRI (C-MRI) across a spectrum of routine neuroradiologic examinations. Materials and Methods: In this single-center retrospective study, 26 patients underwent imaging with a commercially available, FDA-cleared deep learning-accelerated MRI reconstruction algorithm (Deep Resolve, Siemens Healthineers), and C-MRI on a Siemens 3 T MAGNETOM Vida scanner between October 24 and November 14, 2023. A total of 113 sequence pairs were acquired across multiple body parts (brain [n = 28], cervical spine [n = 24], thoracic spine [n = 16], lumbar spine [n = 14], internal auditory canals [n = 5], sella [n = 5], neck [n = 5], temporomandibular joints [n = 6], brachial plexus [n = 4], and orbits [n = 6]) and sequences (T2 [n = 38], T1 [n = 30], short tau inversion recovery [n = 21], T1 post-contrast [n = 17], T2 fluid attenuated inversion recovery [n = 5], and proton density [n = 2]) and evaluated by 4 neuroradiologists blinded to the acquisition method for image quality using a 5-point Likert scale. Acquisition parameters were extracted from Digital Imaging and Communications in Medicine (DICOM) metadata and statistically compared. Rater preferences and interrater reliability were assessed using nonparametric tests and intraclass correlation coefficients. Results: Deep learning reduced mean scan time by 51.6% (95% CI: 45.7%–57.7%; from 110.8 seconds to 53.7 seconds; P < .001). Image quality assessments using a Likert scale showed scores slightly above neutral for signal-to-noise ratio (mean 3.51; 95% CI: 3.44–3.58), structural delineation (mean 3.51, 95% CI: 3.44–3.56), and overall image quality (mean 3.56, 95% CI: 3.49–3.63). However, poor interrater reliability (intraclass correlation [ICC] range: 0.06–0.33) showed that the observed differences were not consistent, indicating functional equivalence between conventional and deep learning images. Conclusion: Deep learning MRI enabled substantial scan time reductions while maintaining image quality.