AI in radiology: powerful tools, persistent blind spots

Artificial intelligence has moved from the margins of radiology into the machinery of everyday practice, and a new narrative review published in BMC Artificial Intelligence argues that the technology’s future depends less on raw accuracy than on interpretability, fairness, and the willingness of radiologists to stay in charge. The review, led by Amy Avakian and Garrett Barfoot and published on 14 January 2026, synthesizes literature from 2018 to 2025 drawn from PubMed, Scopus, and Google Scholar, focusing on convolutional neural networks, vision-language models, workflow optimization, bias, and regulation under the US Food and Drug Administration’s Software as a Medical Device framework.

The review’s central finding is that AI tools are already reshaping diagnostic radiology through triage, image classification, report generation, and administrative automation. Representative studies summarized in the review report area under the receiver operating characteristic curve values ranging from 0.84 to 0.96 for key diagnostic models across tasks such as hemorrhage detection, mammography, and chest X-ray triage. In acute care settings, deep learning triage tools flag emergent pathologies such as intracranial hemorrhage and pulmonary embolism within seconds of image acquisition, dynamically reconfiguring radiologist worklists in real time. The authors link these implementations to reduced time-to-treatment, decreased hospital stay durations, and improved outcomes in critical care settings such as stroke and trauma.

Convolutional neural networks remain the computational backbone of image analysis. These architectures process visual data through successive layers that extract increasingly abstract features, detecting everything from microcalcifications in breast tissue to subtle parenchymal abnormalities and cerebral hemorrhages. Rather than being explicitly programmed, they learn by associating patterns in pixel intensity distributions with labeled diagnostic outcomes. Transfer learning, in which a model trained on a general-purpose dataset such as ImageNet is repurposed for medical imaging, has been particularly valuable because annotated radiology datasets are limited by privacy concerns, labeling costs, and the complexity of expert consensus. The review notes that reported AUCs for CNN-based detection tasks typically exceed 0.90 in large multicenter evaluations, and that large-scale evaluations of AI systems for screening mammography demonstrate the potential to reduce both false positives and false negatives in real-world practice.

Yet the review is emphatic that performance on curated datasets does not guarantee performance in clinics. Models that excel at academic medical centers often underperform in community hospital settings where imaging protocols, scanner types, and patient populations differ significantly. This lack of robustness poses a real barrier to clinical adoption, because models may misclassify or degrade silently in environments that differ even modestly from their training distribution. The authors call for systematic external validation, transparent performance reporting, and continuous model monitoring in real-world environments, noting that FDA-cleared systems such as Aidoc and Arterys Cardio AI have demonstrated measurable reductions in report turnaround times and triage delays, illustrating the transition from experimental validation to real-world deployment.

Vision-language models represent the review’s second major theme. These models fuse transformer-based architectures with convolutional vision backbones to map images to natural language representations. Tools such as BioViL-T and CheXagent can generate free-text radiology reports, perform zero-shot classification, and respond to clinical questions embedded in medical imaging contexts. The review describes their potential to streamline communication between clinicians, improve reporting efficiency, and facilitate real-time decision-making, for example by identifying a lung nodule and generating a coherent sentence describing its size, location, and suggested differential diagnoses.

But the review identifies demographic bias as one of the most pressing concerns for these models. VLMs demonstrated reduced diagnostic accuracy when interpreting images from underrepresented groups, including Black and female patients. The authors attribute this discrepancy to the homogeneity of training datasets, arguing that when models are primarily exposed to data from one demographic group, their performance skews toward that group, resulting in an unequal distribution of diagnostic accuracy across populations. In radiology, where early detection and precision are critical, even slight variations in model performance across groups can compound existing healthcare disparities. The review also flags shallow clinical reasoning as a limitation: while VLMs are proficient at mimicking the style and structure of radiology reports, they often lack the contextual awareness needed for nuanced interpretation, and a model might identify bilateral opacities and report them as consistent with pulmonary edema without integrating laboratory values, comorbid conditions, or hemodynamic status.

Beyond image interpretation, the review documents a quieter revolution in administrative intelligence. Machine learning algorithms now predict appointment cancellations and no-shows by analyzing patterns in patient demographics, scheduling history, and external factors such as weather or transportation availability, helping administrative teams proactively adjust imaging schedules. Worklist optimization systems dynamically triage studies based on urgency, modality, and complexity, ensuring that critical cases such as suspected hemorrhages or pulmonary emboli are interpreted first. Predictive tools also forecast daily and weekly imaging volumes across different modalities, helping allocate staff and prevent bottlenecks, while natural language processing models assist in report generation, billing code assignment, and compliance documentation.

The review’s ethical analysis is blunt. Traditional methods of data anonymization are rapidly losing efficacy as re-identification algorithms become increasingly sophisticated, with AI systems capable of reconstructing patient identities from fragmented metadata. Federated learning and synthetic data generation offer promising mitigation by reducing the need for centralized data pooling, but the authors note that these approaches demand regulatory infrastructure, legal clarity, and ethical governance that many institutions currently lack. Interpretability remains urgent: models that deliver accurate predictions but fail to offer transparent reasoning introduce epistemic opacity into patient care, eroding clinical trust and impeding accountability. The review also highlights that many commercially available AI tools are evaluated under the FDA’s Software as a Medical Device framework, including 510(k) clearances and De Novo classifications, while ongoing challenges remain around post-market surveillance, algorithm updates, transparency of training data, and alignment with international regulations.

Open-source initiatives and benchmark datasets such as RSNA CXR, MIMIC-CXR, and The Cancer Imaging Archive are described as essential for reproducibility, fairness auditing, and transparent benchmarking, enabling independent validation and supporting regulatory trustworthiness. The authors argue that successful implementation requires not only algorithmic accuracy but also interpretability, fairness, and ethical infrastructure, and that radiologists must play a central role in validating and supervising these technologies to ensure they support patient-centered, equitable care. They call for future research to prioritize multimodal learning that integrates imaging, genomic, and clinical data while emphasizing transparency, prospective validation, and adaptive regulation, with collaborative consortia using federated frameworks to evaluate model drift and ensure equitable generalization.

The review concludes that AI is transforming radiology into a hybrid of machine-driven augmentation and clinical oversight, and that the direction of this transformation will be shaped by practicing radiologists who understand the clinical stakes. The authors frame the technology not as a replacement for human expertise but as a force that radiologists must guide, critique, and adapt, warning that the tools built will only be as ethical, equitable, and effective as the frameworks they are embedded within. As AI tools increasingly enter clinical workflows, the review’s message is that understanding their performance, limitations, and barriers to adoption is no longer optional; it is central to safe and ethical practice.

Subject of Research: A narrative review of artificial intelligence methods, clinical impact, and ethical challenges in diagnostic radiology

Article Title: Artificial intelligence in radiology: a narrative review of current methods, clinical impact, and future directions

Article References: Avakian, A., & Barfoot, G. (2026). Artificial intelligence in radiology: a narrative review of current methods, clinical impact, and future directions. BMC Artificial Intelligence, 2(1), Article 1. https://doi.org/10.1186/s44398-025-00020-7

Image Credits: AI Generated

DOI: 10.1186/s44398-025-00020-7

Keywords: artificial intelligence, radiology, convolutional neural networks, vision-language models, medical imaging, demographic bias, interpretability, FDA regulation, federated learning, workflow optimization, mammography, clinical validation