Chong Hyun Suh

Plain English
This study tested whether a cerebral embolic protection device—a filter placed in arteries leading to the brain—reduces new brain lesions during transcatheter aortic valve replacement (TAVR), a procedure that often sends small clots to the brain. Patients who received the device had a median of 1 new brain lesion versus 6 in the unprotected group, and the total volume of lesions was also much smaller. While the stroke rate at 30 days was similar between groups, the dramatic reduction in silent brain injury warrants further study in larger trials.

Plain English
In a study of five radiologists reviewing chest X-rays with AI-generated draft reports, reading time dropped from about 26 seconds to 19 seconds per scan over seven rounds, and acceptance of unedited AI reports rose from 55% to 60%. Quality and agreement with the AI drafts were stable for normal scans but varied for abnormal ones. The results show AI can speed up radiology work, but human oversight remains essential for complex or abnormal cases.

Plain English
Researchers built and tested a two-step AI system that automatically measures brainstem regions on routine brain MRI scans and then uses those measurements to distinguish between three Parkinsonian diseases: Parkinson's disease, progressive supranuclear palsy, and multiple system atrophy. The system achieved diagnostic accuracy above 0.91 in both internal and external test groups, and a radiology resident performed better when assisted by the tool. This could reduce the time and expertise needed to classify these difficult-to-distinguish conditions.

Plain English
This review examines how imaging tools such as DXA, CT, MRI, and bioelectrical impedance can measure body fat more precisely than BMI or waist size alone. CT and MRI offer the most accurate fat measurements, particularly for harmful fat stored around internal organs, while DXA is more accessible for large studies. Regulators are increasingly accepting these imaging measures as valid endpoints in clinical trials of weight-loss drugs, improving how treatment effects are captured.

Plain English
Researchers tested whether GPT-4o and o1 could automatically check whether published medical AI studies followed a reporting checklist. Both models performed well on clear, factual items—such as whether a model's name and version were stated—but struggled with items requiring interpretation of context, such as whether test data were truly independent. The results suggest AI can help automate quality checks for straightforward reporting criteria but still needs human review for nuanced judgments.

Plain English
This study tested four AI language models on 401 neuroradiology quiz cases and compared their accuracy to trained neuroradiologists. The best AI model identified the correct diagnosis among its top three answers only 49% of the time, while radiologists achieved 68–80%. The AI also frequently missed key imaging findings, produced acceptable reasoning in fewer than one in three cases, and performed worse when no written clinical history was provided.

Plain English
An international panel of 57 experts developed REFINE, a 44-item checklist for reporting studies that use large language models and foundation models in medical research, including imaging AI. The checklist covers six areas—from how models are identified and accessed to how data and results are reported—and is supported by an online platform. Adopting REFINE should make AI studies in medicine more transparent and easier to compare across institutions.

Plain English
This meta-analysis compared two MRI techniques for detecting tiny brain bleeds called cerebral microbleeds: susceptibility-weighted imaging (SWI) and gradient-recalled echo (GRE). SWI found roughly 1.6 times as many microbleeds as GRE, with the advantage growing when thinner imaging slices were used. This matters because safety thresholds for Alzheimer's drug trials are currently based on GRE, so different limits may be needed when SWI is used instead.

Plain English
This nationwide Korean study compared 368 patients with CADASIL—a genetic form of small vessel brain disease—by sex, finding that men had a much higher rate of ischemic stroke and experienced their first stroke about six years earlier than women. Men also had more small brain lesions called lacunes, while women tended toward greater white matter damage. The findings show that sex shapes how this genetic disease plays out and should inform how patients are monitored and treated.

Plain English
Researchers reviewed 246 studies that tested large language models in radiology and checked whether they followed key reporting standards. Most studies failed to document basic details like model version, access date, or the exact prompts used, and these gaps persisted even in more recent publications. Without this information, other researchers cannot reliably reproduce or compare results, undermining the trustworthiness of the field.

Plain English
New Alzheimer's treatments—lecanemab and donanemab—work by clearing amyloid plaques from the brain but can cause a side effect called ARIA, which shows up as brain swelling or microbleeds on MRI. This review covers what ARIA looks like on imaging, how common it is, which patients are at highest risk, and how radiologists should grade and manage it. As these drugs enter clinical use, radiologists play a central role in safe patient monitoring.

Plain English
Researchers evaluated how well published studies on AI language models in top medical journals followed the MI-CLEAR-LLM reporting checklist. While nearly all studies named the model, fewer than 20% documented how they managed randomness in AI outputs, and only 13% confirmed that test data were independent from training data. These widespread gaps mean most published AI accuracy claims in medicine cannot be fully verified or reproduced.

Plain English
Researchers compared four blood-based inflammation scores in 156 patients with polymyalgia rheumatica versus 408 with rheumatoid arthritis to see if these scores could help tell the two conditions apart. The C-reactive protein-to-albumin ratio performed best, with an area under the curve of 0.82, and combining all four scores improved accuracy slightly further. These simple, inexpensive lab ratios could serve as useful diagnostic aids when distinguishing polymyalgia rheumatica from rheumatoid arthritis.

Plain English
A systematic review compiled published evidence on which imaging tests are used to find the cause of pulsatile tinnitus—a rhythmic sound in the ear. MRA (magnetic resonance angiography) identified most conditions that traditionally required catheter-based angiography, but few studies directly compared imaging methods head-to-head. The lack of standardized protocols means imaging choices are still largely driven by physician preference, and prospective comparison studies are needed to establish evidence-based guidelines.

Plain English
This meta-analysis pooled data from nine phase 3 trials of amyloid-clearing antibodies for Alzheimer's disease to quantify the rate of brain imaging side effects called ARIA. After accounting for heterogeneity, about 25% of treated patients developed ARIA with edema and 18% developed microbleeds, with patients carrying two copies of the APOE ε4 gene facing a 5.6-fold higher risk of the edema type. These rates and risk factors are essential for patient selection and safety monitoring in clinical practice.

Plain English
A domain-specific AI model trained on nearly 9 million chest X-ray reports was evaluated for its ability to generate preliminary radiology reports. Radiologists accepted the AI-generated reports without changes 70% of the time—nearly matching each other at 73%—and ranked them highest in quality among AI and general-purpose model outputs. The model showed particularly high sensitivity for critical findings like pneumothorax, suggesting it can meaningfully assist radiologists in routine chest X-ray interpretation.

Plain English
This is a published correction to an earlier review article on amyloid-related imaging abnormalities and anti-amyloid therapy for Alzheimer's disease.

Plain English
Using a large Korean registry of 1,832 rheumatoid arthritis patients on biologic drugs, researchers found that infusion or injection reactions occurred in about 10% of patients and were more common in younger patients, those with secondary Sjogren's syndrome, and those who had previously used leflunomide. Abatacept, tocilizumab, and golimumab carried significantly lower reaction risk than infliximab. After a reaction, prescribing patterns shifted away from TNF inhibitors toward tocilizumab and JAK inhibitors.

Plain English
This article reviews the role of imaging in monitoring patients receiving CAR T-cell therapy for cancer, a rapidly expanding treatment category. It covers how tumor response is assessed differently depending on cancer type, how imaging detects serious side effects like cytokine release syndrome and brain toxicity, and how the Lugano criteria provide a standardized framework for lymphoma response evaluation. Radiologists need familiarity with these specifics to support CAR T-cell therapy programs effectively.

Plain English
Among 516 elderly Korean patients with rheumatoid arthritis on advanced biologic or targeted therapies, 10.5% met criteria for difficult-to-treat disease—meaning they could not achieve low disease activity despite multiple drug trials. These patients had persistently higher disease activity scores at one and two years, and factors like higher symptom burden and fewer prior standard drugs predicted this pattern. The findings highlight a meaningful subgroup of elderly RA patients who need closer monitoring and more creative treatment approaches.

Plain English
This study identified genetic and imaging differences between CADASIL patients who had experienced ischemic stroke versus those who had not. Patients with stroke had more lacunes and worse white matter disease on MRI, and were more likely to carry a variant in exon 3 of the NOTCH3 gene, while exon 11 variants were associated with a lower stroke risk. Knowing which gene variant a CADASIL patient carries may help predict their individual stroke risk.

Plain English
A phase 2b clinical trial tested three doses of epaminurad—a new drug that blocks uric acid reabsorption in the kidneys—against placebo in 169 gout patients over 12 weeks. All three doses significantly reduced blood uric acid below the treatment target, with the 9 mg dose achieving the target in 89% of patients, comparable to the established drug febuxostat. Epaminurad was well tolerated with no kidney or liver safety signals, supporting its advancement to larger trials.

Plain English
An AI model was developed and validated to automatically score medial temporal lobe atrophy on brain MRI—a key indicator of Alzheimer's-related memory decline—in patients with cognitive impairment. The model achieved 82–87% accuracy on internal test cases and 82–85% on an external dataset, matching the performance of a machine learning comparison model. Automated scoring of this visual rating scale could make dementia assessment faster and more consistent in clinical practice.

Plain English
A systematic review and meta-analysis pooled safety data from studies of gadolinium contrast use in children. Acute adverse reactions occurred in fewer than 0.3% of cases, no children developed nephrogenic systemic fibrosis, and only linear gadolinium agents—not macrocyclic ones—showed measurable accumulation in the brain with repeated use. These findings confirm that gadolinium contrast is generally safe in pediatric patients, with macrocyclic agents preferred to minimize any brain deposition.

Plain English
Researchers used brain connectivity maps derived from pre-surgery MRI scans to predict which Parkinson's disease patients would benefit from deep brain stimulation targeting the globus pallidus for freezing of gait. Adding connectivity features between the stimulated brain region and areas involved in motor planning improved prediction accuracy from 65% to 77% compared to using patient demographics alone. These results support using pre-operative brain connectivity data to personalize deep brain stimulation programming.

Plain English
Brain MRI scans from 745 adults aged 60 and older were analyzed to see whether hearing loss was linked to brain volume loss. Even mild hearing loss was associated with reduced total gray matter, and moderate-to-severe hearing loss was specifically linked to hippocampal shrinkage—a region critical for memory. The findings suggest that treating hearing loss early may help protect the brain from age-related volume loss and reduce dementia risk.

Plain English
This retrospective study followed 120 patients with rheumatoid arthritis-associated lung disease (RA-ILD) over an average of four years. Lung disease progressed in 58% of patients and the mortality rate was 21%, with the scarring lung pattern (usual interstitial pneumonia) and overall lung disease duration being the strongest predictors of death. The high progression and mortality rates underscore the need for regular lung monitoring and individualized treatment in RA patients with lung involvement.

Plain English
Seven AI language models were tested against four human readers on 71 pediatric radiology cases to see how well they could flag cases needing immediate attention. The best-performing AI, Gemini Flash, achieved 83% accuracy overall—comparable to the most accurate human reader—but showed poor sensitivity for the most critical finding category, detecting only 3 of 11 such cases. AI may help prioritize urgent pediatric radiology cases on a worklist, but unreliable detection of the most critical cases limits its standalone use.

Plain English
Researchers synthesized two new compounds—ETI41 and ETI60—that selectively block a set of immune receptors inside cells (endosomal Toll-like receptors) involved in triggering autoimmune flares. Both compounds reduced inflammation in cell and animal models, and oral administration improved symptoms in mouse models of psoriasis and lupus without affecting surface immune receptors. These compounds are promising drug candidates for treating autoimmune diseases driven by aberrant endosomal TLR activation.

Plain English
This large retrospective study examined whether gadolinium contrast used in MRI raises the risk of acute kidney injury in hospitalized adults, comparing over 35,000 scans with and without contrast after accounting for patient differences. Patients who received contrast actually had a lower rate of acute kidney injury, and this held true even in patients with reduced kidney function. The results suggest gadolinium-based contrast agents are safe for the kidneys across a wide range of renal function.

Plain English
A six-year review of 23,429 iodine contrast injections in children found that allergic-like reactions occurred in 0.89% of cases and physiologic reactions in 0.55%, with the vast majority being mild. Older children and those with a prior reaction history were at highest risk, and switching to a different contrast agent at re-exposure roughly halved the recurrence rate. These data provide pediatric-specific risk estimates that can guide pre-procedure screening and contrast agent selection in children.

Plain English
This real-world Korean study compared outcomes in 179 lupus nephritis patients who did or did not receive tacrolimus as maintenance therapy. Complete remission rates were lower in the tacrolimus group at one and two years, though overall response rates and rates of serious complications were similar between groups. Tacrolimus appears to be a viable—if not superior—maintenance option for patients who have not achieved remission on standard therapy.

Plain English
Postmenopausal women with rheumatoid arthritis were divided into those with and without bone erosions on X-ray, and their hip bone density was measured in three dimensions using a specialized DXA-based tool. Women with erosions had significantly lower bone density across all three bone compartments at the femoral neck, and higher disease activity independently predicted worse cortical bone density. Keeping rheumatoid arthritis disease activity low is essential for protecting bone structure, particularly in postmenopausal women.

Plain English
This study compared cardiovascular risk scores and tests for silent atherosclerosis in 67 Korean women with lupus against 37 healthy controls. Standard risk scores underestimated cardiovascular risk in lupus patients compared to a Korea-specific calculator, and ultrasound measures of artery wall thickness correlated with all risk scores. Using a country-appropriate cardiovascular risk tool improves detection of hidden heart disease risk in Korean lupus patients.

Plain English
Researchers measured the thickness of the temporal muscle on brain MRI scans—a marker of overall muscle mass—in 285 patients with the most aggressive form of brain cancer, glioblastoma. Patients with thinner temporal muscles (indicating poor nutritional and physical status) survived significantly less time, and this effect grew stronger with advancing age. Temporal muscle thickness is a simple measurement from routine scans that provides independent survival information beyond standard clinical and molecular factors.

Plain English
This article presents an updated version of the MI-CLEAR-LLM checklist, a reporting standard for studies that test AI language models in healthcare settings. The revised checklist adds guidance for models accessed through programming interfaces and open-source systems, covering how models are identified, accessed, adapted, and run, as well as how randomness in outputs is managed and whether test data are truly independent. Following this checklist helps ensure AI medical studies are transparent enough to be reliably evaluated and replicated.

Plain English
Machine learning models were trained to distinguish two conditions that look similar on brain MRI—idiopathic normal pressure hydrocephalus and progressive supranuclear palsy—using automated measurements of brain volume and shape. A support vector machine using volumetric features achieved near-perfect accuracy (area under the curve of 0.98), with brainstem and fluid space measurements being the most informative. Automated brain volumetry can reliably separate these two diagnoses without requiring subjective visual assessment.

Plain English
Six AI language models were tested on 56 complex neurology cases from top medical journals and compared to three neuroradiologists. The best model, Claude 3.5, reached 80% accuracy when given both images and text, matching performance roughly equivalent to first-year radiology fellows, though image-only accuracy was much lower for all models. The findings confirm AI can handle text-heavy diagnostic reasoning but still struggles with independent image interpretation.

Plain English
This study compared 70 stroke patients with CADASIL—a hereditary small vessel brain disease—to 70 matched patients with non-genetic small vessel strokes over a 10-year follow-up. CADASIL patients had strokes in different brain locations, more often involving multiple simultaneous lesions, and a 50% recurrence rate with a 15-fold higher risk of repeat stroke in multivariable analysis. These distinct patterns and risks argue for tailored monitoring and management strategies in CADASIL patients after a first stroke.

Plain English
Researchers trained and validated an AI model that simultaneously locates and grades white matter hyperintensities—bright spots on brain MRI linked to small vessel disease—in patients with memory and thinking problems. The model achieved high accuracy on both internal and external test sets, closely matching the ratings of experienced neuroradiologists. Automated grading of these lesions could help clinicians consistently assess and monitor brain vascular damage at scale.

Plain English
This article traces the evolution of the Korean radiology board exam from paper-based memory tests to tablet-based, case-driven problem-solving with real-world imaging formats including stacked and video images. The authors argue that exams must continue to evolve—moving beyond arrow-marked images toward independent lesion identification—and must incorporate AI-related competencies. Maintaining validity and reliability throughout this transition is the central challenge for the examination committee.

Plain English
Researchers developed an AI that generates synthetic contrast-enhanced brain MRI images from standard non-contrast scans, eliminating the need for gadolinium injections to visualize the choroid plexus—a brain structure linked to neurodegeneration. The synthetic images closely matched real contrast-enhanced scans and allowed accurate measurement of choroid plexus size, revealing that patients with atypical Parkinsonism had larger choroid plexuses than those with standard Parkinson's disease. This approach could enable safer, repeated monitoring of brain changes without exposing patients to contrast agents.

Plain English
This pilot study measured blood levels of two fat-derived signaling molecules—ceramide 16 and sphingosine-1-phosphate—in women with lupus, with and without kidney involvement. The ratio of ceramide to sphingosine-1-phosphate was significantly higher in lupus patients overall and even higher in those with kidney disease, and it correlated with established markers of lupus activity. This ratio may serve as an accessible blood test to track disease severity and kidney involvement in lupus.

Plain English
Two new drugs—HCB-5300 and HCB-5400—were tested in mice with chemically induced colitis to see if blocking a protein called STAT3 could reduce gut inflammation. Both drugs reduced weight loss, intestinal damage, and levels of an inflammatory protein called IL-6, with HCB-5400 showing stronger effects across most measures. These results support further development of HCB-5400 as a potential treatment for inflammatory bowel disease.