RADIOLOGY DEPARTMENTS ARE FACING A CRISIS that has nothing to do with technology and everything to do with arithmetic. The global volume of medical images is growing at approximately 15 percent annually, driven by ageing populations, expanding cancer screening programmes, and the proliferation of imaging as a primary diagnostic modality. The global pool of trained radiologists is growing at approximately 2 percent annually. The mathematical consequence of this divergence is already manifest in clinical practice: radiologists in major hospital systems routinely read 80 to 100 studies per shift, under time pressure that introduces fatigue-related error, and reporting backlogs in many NHS trusts and Indian public hospitals reach weeks or months, during which cancers grow and fractures go unset. Artificial intelligence — specifically deep learning systems trained on millions of annotated medical images — is the only available technology capable of addressing this gap at the required scale and speed.
By late 2025, the AI-enabled medical device market had reached $13.7 billion, with projections to exceed $255 billion by 2033. The FDA had cleared 1,451 AI-enabled medical devices cumulatively, of which radiology imaging applications represent 76 percent — the single largest category. As of 2025, 54 percent of US hospitals with more than 100 beds report using AI in radiology, primarily for image interpretation (82 percent) and worklist prioritisation (48 percent). These tools have demonstrably reduced average report turnaround times from 11.2 days to as low as 2.7 days in adopting institutions. AI tools have also been shown to reduce radiologists’ overall workload by up to 53 percent, according to a 2025 systematised review in Health and Technology.
The Clinical Evidence Base
The clinical evidence for AI imaging performance has matured substantially in the past two to three years, moving from proof-of-concept studies on retrospective datasets to prospective, multi-centre, real-world validation. AI-assisted fracture detection has become one of the most validated clinical applications. A multi-reader, multi-case study of 340 radiographic exams evaluated an AI fracture detection tool against non-specialist radiologists: with AI assistance, sensitivity increased from 72 to 80 percent, specificity from 81 to 85 percent, missed fractures decreased by 29 percent, and false positives by 21 percent — without adding reading time. The greatest improvements were in detecting non-obvious, cortically-impacted fractures that represent the highest-risk category of diagnostic miss.
In oncology, the evidence is equally compelling. A landmark international, paired, non-inferiority confirmatory observational study — the PANORAMA study — published in The Lancet Oncology in January 2026 evaluated AI performance in detecting pancreatic cancer on standard-of-care CT scans, comparing AI to radiologist reads. AI-driven disease detection systems for various cancers have demonstrated accuracy of 81 to 95 percent in identifying infections and lesions 2 to 3 weeks before symptom emergence. In breast cancer screening, AI-assisted reading through human-AI delegation strategies has been shown to reduce screening costs by 17.5 to 30.1 percent compared with radiologist-only approaches, while maintaining or improving detection sensitivity. In lung cancer screening, AI analysis of CT scans has identified incidental pulmonary findings — early-stage nodules, ground-glass opacities — at rates substantially higher than human-only reading in high-volume screening programmes.
Digital Pathology: The Next Frontier
While radiology AI has received the greatest clinical and commercial attention, the transformation of pathology — the discipline responsible for the definitive diagnosis of cancer through microscopic examination of tissue — may prove equally significant. Pathology has historically been resistant to automation: the diagnosis of cancer from a histological slide requires the integrated assessment of cell morphology, tissue architecture, spatial relationships, and molecular context that has been considered uniquely resistant to algorithmic reproduction. Deep learning systems trained on digitised whole-slide images are now demonstrating that this assumption was wrong.
Convolutional neural networks applied to digital pathology slides have demonstrated clinical-grade accuracy for cancer grading, cell segmentation, biomarker identification, and even the prediction of molecular phenotype — the genetic subtype of a cancer — directly from the tissue image, without the need for expensive molecular testing. Multimodal AI models integrating radiological imaging with digital pathology and molecular data have shown promise in improving risk stratification beyond traditional staging systems. A multimodal AI system integrating imaging, clinical, and molecular data for predicting breast cancer recurrence was presented at the 2025 San Antonio Breast Cancer Symposium, demonstrating superior predictive accuracy compared with the TAILORx genomic test that currently guides treatment decisions for hundreds of thousands of patients annually.
Foundation Models: The Architecture of Scale
The most significant recent development in medical imaging AI is the emergence of foundation models — large, general-purpose AI systems pre-trained on enormous and diverse datasets that can be fine-tuned for specific clinical tasks with relatively small amounts of labelled data. Aidoc’s CARE1 foundation model, which received FDA clearance in February 2025 as the first foundation-model-powered clinical AI, demonstrated that AI systems can be built for generalisability across clinical contexts rather than narrow, single-task applications. The ‘AI Cockpit’ concept now emerging in leading radiology departments envisions a dashboard through which every incoming image runs through a suite of AI checks simultaneously — cancer detection, fracture identification, haemorrhage alerting, organ anomaly flagging — with automated tagging by urgency and preliminary measurements, all before a radiologist opens the study. Such a system could reduce the 3 percent of diagnostic errors attributable to missed incidental findings — a proportion that sounds small but translates to tens of thousands of adverse patient outcomes annually in high-volume imaging environments.
India: Radiology AI as a Leapfrog Technology
India faces a severe and worsening radiology workforce crisis. With fewer than 15,000 qualified radiologists for a population of 1.4 billion — compared with approximately 25,000 in the US serving one-quarter the population — the per-capita deficit is extreme, particularly outside major metropolitan centres. AI imaging tools that can function as competent first readers, prioritise urgent studies, and flag findings for limited specialist review represent a genuine leapfrog technology for India’s healthcare system. Platforms including Qure.ai, a Bengaluru-based company, have developed and validated AI tools for chest X-ray interpretation, CT brain analysis, and bone age assessment specifically calibrated for Indian patient populations and deployed at scale in Indian public health programmes. Qure.ai’s tools have been validated in real-world deployments in India’s National TB Elimination Programme, demonstrating that AI imaging can function effectively in resource-limited, high-volume public health contexts. For Smart Labtech and NSH’s laboratory instruments and diagnostics readership, the implications extend to laboratory-based imaging as well: AI-powered analysis of digital microscopy images, flow cytometry data, and haematology slides is rapidly advancing, extending the AI imaging revolution from radiology suites into the diagnostic laboratory.
– Sunil
