Journal of the American College of Radiology : JACR最新文献

Large language models (LLMs) are increasingly being explored for a wide range of applications in radiology, offering the potential to enhance clinical workflows, improve diagnostic accuracy, and support patient communication. In this scoping review the authors examine the current and emerging uses of LLMs on radiology text, focusing on areas such as report generation, structured data extraction, workflow optimization, and clinical decision support. A literature search was conducted on PubMed and Embase, and a total of 69 articles were included in the review. The capabilities and limitations of existing approaches were assessed, and key methodologic considerations were discussed, including transparency and bias, while identifying critical gaps in validation and generalizability. Overall, LLMs demonstrated strong performance in workflows such as report simplification and translation but produced mixed results in classification tasks. Certain methods such as fine-tuning and structured prompt generation improved LLM accuracy. In assessing the characteristics of the included studies, although most studies performed well in documenting the independence of their testing and training datasets and LLM prompting methods, fewer than half of studies explicitly attempted to manage the inherent stochasticity of LLMs. By synthesizing recent advancements and outlining future directions, the aim of this review was to guide clinicians, researchers, and health care stakeholders in responsibly harnessing the transformative potential of LLMs in radiologic care.

Purpose: Lower socioeconomic status (SES) and longer distance to providers have been linked to disparities in health care access, often delaying diagnostic evaluation of acute conditions. Determining the effects of nonmedical factors on the acuity of MRI findings is critical for addressing health care inequities and improving access to medical imaging. The goal of this study was to characterize the effects of neighborhood SES and distance from home address to imaging center on acuity of brain MRI findings.

Methods: The investigators evaluated brain MRI scans acquired from March 2019 to March 2020. Neighborhood SES was quantified using the area deprivation index (ADI), and acuity was categorized by board-certified neuroradiologists (1, 2, or 3, with 3 indicating the highest acuity). Distance between home address and imaging center was calculated in miles and log₂ transformed. The χ² test and analysis of variance were used to assess groupwise differences in acuity. A multinomial baseline-category logit model was used to assess the relationship between z-scored ADI and log distance on acuity, controlling for age, sex, race, marital status, insurance type, and encounter type.

Results: The final cohort consisted of brain MRI scans from 4,813 individuals (mean age, 58.5 ± 17.3 years; 2,733 women). Acuity varied significantly across sex, race, marital status, insurance type, encounter type, and imaging center. Regression analysis revealed a 1-SD increase in z-scored ADI (ie, lower neighborhood SES) resulted in significantly higher acuity of inpatient brain MRI scans (acuity 2 > 1 odds ratio [OR], 1.34 [P = .0016]; acuity 3 > 1 OR, 1.32 [P = .0012]) and emergency department scans (acuity 2 > 1 OR, 1.27; P = .045). A twofold increase in distance resulted in significantly higher acuity of brain MRI findings for all scans (acuity 2 > 1 OR, 1.07 [P = .026]; acuity 3 > 1 OR, 1.15 [P < .001]).

Conclusions: Lower neighborhood SES for inpatient and emergency department scans and greater distance between home address and imaging center for all scans result in significantly higher odds of more acute brain MRI findings.

Purpose: Industry-sponsored research payments (hereafter research payments) play an important role in the advancement of the medical sciences. Recent analysis of the Open Payments program (OPP) demonstrated that research payments are increasingly dispersed through noncovered entities (NCEs) across specialties. It is unknown how research payments are distributed to radiologists. The aims of this study were to assess trends in research payments to radiologists between 2015 and 2024 and to characterize payments distributed through NCEs.

Methods: A retrospective analysis of research payments in the OPP was performed, and payments accepted by radiologist principal investigators at NCEs were analyzed. Linear regression and descriptive statistics were used to examine annual payment trends.

Results: Between 2015 and 2024, 57,772 research payments totaling $824.2 million were accepted by 2,963 radiologists. Annual payments increased 50%, driven by a 44% increase in payments to NCEs (P < .01). At NCEs, median payments per principal investigator increased from $13,718 (interquartile range: $3,623-$53,795) to $24,350 (interquartile range: $5,500-$80,720) (P < .001). In 2024, male radiologists received 88% ($58.8 million) of NCE payments, a 54% increase (P < .01). Five manufacturers accounted for 40.4% of total research payments to radiologists.

Conclusions: Research payments to radiologists have increased significantly because of increases in payments to NCEs, concentrated among a select group of radiologists. When transparent and appropriately managed, NCEs represent valuable partnerships improving patient care, yet further research remains imperative to understand changes in research funding pathways.

Objective: The purpose of this study is to examine the performance of Chat Generative Pre-trained Transformer (GPT)-4vision (GPT-4v) and GPT-4omni (GPT-4o) on the ACR's Diagnostic Radiology in-Training (DXIT) examination, comparing performance on image-based and text-only questions.

Methods: In all, 1,136 publicly available DXIT examination questions were input into GPT-4v and GPT-4o with a prompt asking the large language model to provide its answer, rationale, and confidence level (0-100). Accuracy of each model across different categories was then analyzed, with χ² tests to compare proportions, t tests to compare means, and receiver operating characteristic curves to evaluate confidence levels.

Results: GPT-4o and GPT-4v achieved accuracies of 73.5% and 69.3%, respectively (P < .0001) while scoring 55.6% and 50.3% on image-based questions (P < .0001). Receiver operating characteristic curves of confidence levels and correctness produced areas under the curve of 0.64 and 0.66 for GPT-4o and GPT-4v, respectively.

Discussion: GPT-4o outperformed GPT-4v on nearly every metric, with both models outperforming the national average performance of postgraduate year 3 radiology residents (61.9%) on the 2022 DXIT examination. However, performance on image-based questions remains significantly worse than text-only questions, and both models score below radiology trainees from the same cohort. Both models exhibit limited ability to predict correctness using an intrinsic confidence level. Use of ChatGPT for test preparation and image interpretation must therefore be approached with caution.

Purpose: To apply a Quality and Safety Continuous Process Improvement approach guided by Continuous Quality Improvement and Plan-Do-Study-Act (PDSA) cycles to develop, refine, and assess a digital reminder program's effect on Screening Mammography Missed Care Opportunity (SM-MCO) rates.

Methods: Study conducted at two Federally Qualified Community Health Centers and a mobile mammography unit. The pre-PDSA period was October 2020 to June 2023, and the post-PDSA period was July 2023 to January 2025. PDSA 1 launched a multilingual Short Messaging System (SMS) reminder across all sites. PDSA 2 standardized reminder process. PDSA 3 implemented a SM educational video. The primary outcome assessed the PDSA cycles' effect on SM-MCO rates. The secondary outcome assessed digital engagement. Quality improvement Statistical Process Control p-chart tracked appointment-level data. Univariate and logistic regression analyses assessed primary and secondary outcomes.

Results: In all, 18,654 appointments were included in the analysis; average age was 56.8 (SD = 9.6 years), and 51.9% identified as Hispanic. The overall SM-MCO rate declined from 29.2% pre-PDSA to 26.9% post-PDSA (P < .001). Appointments with SMS had a 35% SM-MCO rate, compared with 21.7% without (P < .001). Appointments with digital engagement had an SM-MCO rate of 21.7% compared with 40.4% without engagement (P < .001). Appointments that received and viewed the video had an SM-MCO rate of 11.5% compared with 26.9% without it (P < .001).

Conclusion: Although a modest decrease in overall SM-MCOs rate was observed, SM-MCO rates were higher among appointments that received SMS reminders but lower among appointments with digital engagement, underscoring the digital divide complexity. Quality Improvement frameworks can continuously monitor and refine digital strategies to increase access to radiology.

Objective: In the 26 years since its establishment, the Holman Research Pathway (HRP) has changed significantly. For example, a study published in 2018 found that interest among diagnostic radiology (DR) residents in the program had waned significantly, raising questions about the program's future. In this study, we sought to better understand the effectiveness of the HRP among DR residents, with a focus on the residency research productivity and career outcomes of DR residents who have completed the program.

Methods: We identified DR graduates of the HRP between 2003 and 2023 using the ABR's website and collected data regarding demographics, research output, and career outcomes from publicly available online sources. Research productivity was measured by first-author publications during residency and first- or last-author publications within 30 months after graduating from residency. Journal impact factors, citations, grant support, and open-access status were recorded. National Institutes of Health funding and academic employment were also evaluated.

Results: Thirty-three DR residents completed the HRP from 2003 to 2023 (mean 1.6 per year); 91% of graduates have completed subspecialty fellowships, 67% currently hold academic positions, and 27% have received National Institutes of Health funding. During training, residents published 64 first-author articles (mean 1.9 per resident) in journals with a median impact factor of 4.7, and 67% of these articles were published in open-access journals. In the first 30 months postresidency, graduates published a mean of 1.5 first- and last-author manuscripts in journals with a median impact factor of 3.5. There was a positive correlation between residency and postresidency research productivity (r = 0.5, P < .01).

Discussion: Although HRP participants in DR demonstrate research productivity comparable to radiation oncology graduates, fewer remain in academic positions, and overall participation has remained low. Increased awareness and support for the HRP may help attract more DR residents.

Artificial intelligence applications for radiology workflow have the potential to improve patient- and health system-level outcomes through more efficient and accurate diagnosis and clinical decision making. For a variety of time-intensive steps, numerous types of applications are now available with variable reported measures and degrees of success. The tools we highlight aim to accelerate imaging acquisition, reduce cognitive and manual burden on radiologists and others involved in the care pathway, improve diagnostic accuracy, and shorten the time to clinical action based on imaging results. Most existing studies have focused on intermediate outcomes, such as task duration or time to the next step in care. In this article, we present an examination of artificial intelligence applications across the medical imaging examination workflow, review examples of real-world evidence on these tools, and summarize the relevant performance metrics by application type. Beyond the more immediately acquired measures, to demonstrate benefit to patient health and economic outcomes, a more integrated assessment is necessary, and in an iterative fashion. To evolve beyond early workflow gains, interoperable tools must be tied to measurable downstream impacts, such as reduced disease severity, lower mortality, and shorter hospital stays, although we acknowledge that current empirical evaluations are limited.