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

Objectives: To introduce a vertically integrated model between a health care service provider and technology developer as a learning accelerator to address challenges in developing and delivering artificial intelligence (AI) into health care.

Methods: The Learning Accelerator Framework is built on four core components that focus on improving patient and health care outcomes: an integrated data registry, a continuous technology development stack, adaptive clinical services, and an iterative learning and development loop. Its application is described in one case study to highlight its operational mechanisms throughout the AI life cycle.

Results: The framework has guided the conceptualization, development, implementation, and national delivery of a multistage AI breast cancer screening workflow, progressing from initial clinical validation (thousands) to population-scale implementation (millions of patients). We demonstrate how iterative learning loops were applied using clinical feedback and real-world data monitoring feedback, which resulted in a multistage AI screening workflow that has achieved a significant absolute increase in cancer detection rate (Δ0.99 cancers per 1,000 examinations [95% confidence interval: 0.59-1.42]) and positive predictive value (Δ0.55 cancers per 100 recalls [95% confidence interval: 0.30-1.03]) with equitable benefits across breast density, race, and ethnic subpopulations.

Discussion: The Learning Accelerator Framework represents a departure from traditional approaches by mitigating challenges, inefficiencies, and delays that impede AI translation, offering a model for AI developers and provider systems seeking to accelerate innovation. The breast AI case study demonstrates how instrumental the framework can be for ensuring ongoing AI implementation effectiveness, fostering clinician trust, and ultimately improving operations, patient outcomes and health equity.

Purpose: The aim of this study was to estimate the 2-year incidence of lung cancer diagnosed as a result of radiologist recommendations for chest CT in neck CT and MRI reports.

Methods: A retrospective observational cohort study was conducted, including all patients without histories of lung cancer with recommendations for chest CT in neck CT and MRI reports from June 1, 2021, to May 31, 2022, in a multi-institution health care system. Outcome data were extracted up to December 31, 2024. Two-year lung cancer incidence was estimated using a person-time calculation to acknowledge censoring with confidence intervals based on quasi-likelihood. Odds of fulfillment of the recommended chest CT for pulmonary nodules relative to other pulmonary abnormalities were estimated using logistic regression.

Results: Two hundred seventy-six of 28,707 (1.0%) consecutive neck, brachial plexus, and parathyroid CT and MRI reports in 273 of 22,173 patients (1.2%) (mean age, 62.5 ± 1 years, 52% women) contained recommendations for chest CT in the absence of prior lung cancer diagnoses. The median follow-up time was 34 months (interquartile range, 24-40 months). One patient (estimated 2-year incidence rate, 0.40%; 95% confidence interval, 0.05%-3.55%) was diagnosed with an incidental indolent adenocarcinoma. Recommended CT was performed in 171 of 273 patients (62.6%) and was less likely to be performed for pulmonary nodules than other pulmonary abnormalities (odds ratio, 0.46; 95% confidence interval, 0.27-0.77).

Conclusions: One year of recommendations for chest CT examinations in neck CT and MRI reports across a multi-institution health care system led to the identification of only a single incidental lung cancer, an indolent adenocarcinoma. These results suggest that the frequency of recommendations for chest CT should likely be substantially decreased, but analysis of larger datasets is needed to inform best practices.

Image quality is central to the accurate interpretation of medical imaging, yet it remains inconsistently defined and assessed across clinical practice. To address this, the ACR has developed the Medical Image Quality Assessment System (MIQAS), a standardized, descriptive framework that characterizes image quality based on its alignment with relevant clinical task requirements. This framework will serve as the image quality assessment standard for all relevant ACR programs, including the ACR Accreditation Program, the Reporting and Data Systems programs, Practice Parameters and Technical Standards, and the ACR Learning Network. In this framework, image quality is defined as the degree to which an image approximates an exact representation of its subject in ways that matter for a specific clinical task. Image quality assessments may be quantitative, semiquantitative, or categorical, but should be reproducible and valid. Under this framework, key image quality elements of an imaging examination are individually scored and aggregated into a composite score on a 5-point scale: 0 (out of standard), 1 (nondiagnostic), 2 (limited), 3 (adequate), and 4 (excellent). For "bounded" image quality factors that involve trade-offs with cost or risk-such as radiation dose in CT-the goal is "adequate" image quality. For unbounded factors without such trade-offs-such as positioning or labeling-the goal is "excellent" image quality. Individual scoring systems will be developed under this overarching framework for specific modalities, organ systems, and diagnostic tasks. Once published, each scoring system becomes an ACR-supported standard, updated periodically based on emerging evidence. In this way, the MIQAS framework is designed to unify image quality assessment across ACR programs, guide local quality improvement efforts, and serve as a unified image quality assessment standard for research, education, and technology development.