AI in radiology refers to software, primarily powered by deep learning, that analyzes medical images like X-rays, CT scans, mammograms, and MRIs to help detect diseases, prioritize urgent cases, and extract health information that might otherwise go unnoticed. It’s the largest category of AI in medicine: of the 950 AI and machine learning devices authorized by the FDA through early 2025, 723 (76%) were radiology devices.
How It Differs From Older Technology
Computer-aided detection, or CAD, has existed in radiology since the 1990s. Those early systems used hand-crafted rules to flag suspicious areas on images, particularly in mammography. The results were disappointing. Several large trials found that traditional CAD delivered no benefit at best and actually reduced radiologist accuracy at worst, increasing recall and biopsy rates. Radiologists spent roughly 20% more time per study just dismissing false alarms.
Modern AI is fundamentally different because it runs on deep learning, a technology that learns directly from data rather than following pre-programmed rules. In a well-known benchmark test where software identifies everyday objects in photographs, traditional computer vision methods made five times as many errors as a human. Within a few years, deep learning surpassed human performance on the same test, cutting errors to about half the human rate. That leap in visual perception is what makes today’s radiology AI a genuinely new tool rather than a repackaged version of old CAD.
Traditional CAD was also task-specific: each system had to be painstakingly engineered for one narrow job. Deep learning is task-agnostic. The same underlying architecture can be trained on chest X-rays, brain CTs, or pathology slides. It can also be fine-tuned for new tasks with far less effort, which is why the number of FDA-cleared radiology AI products has grown so quickly.
How the Technology Reads an Image
Most radiology AI uses a type of deep learning architecture called a convolutional neural network, or CNN. A CNN processes a medical image through a series of layers. Early layers detect simple patterns like edges and textures. Each successive layer combines those patterns into increasingly complex features: shapes, structures, and eventually recognizable anatomy or abnormalities. The final layers connect these high-level features to produce a prediction, such as “nodule present” or “no fracture detected.”
These networks are typically trained through supervised learning, meaning they study thousands or millions of images that have already been labeled by radiologists. Over time, the network learns which pixel patterns correspond to which diagnoses. Newer systems go further by combining image-reading networks with large language models, allowing them to generate written descriptions of what they see, essentially drafting portions of a radiology report.
What AI Does in Practice
Radiology AI falls into several practical categories, though the boundaries between them are blurring as the technology matures.
- Detection and diagnosis: AI flags abnormalities that a radiologist then confirms or dismisses. In lung nodule detection on CT scans, AI models achieve sensitivity rates of 86% to 98%, compared with 68% to 76% for radiologists alone. The tradeoff is specificity: AI ranges from 77.5% to 87%, while radiologists hit 87% to 92%. In plain terms, AI catches more real nodules but also flags more false positives. This makes it most useful as a second reader that ensures fewer cancers are missed, with the radiologist filtering out the false alarms.
- Triage and prioritization: Some AI tools scan incoming studies and move urgent cases, like suspected brain bleeds, to the top of a radiologist’s reading list. The goal is faster treatment for critical patients, though real-world results have been mixed. A prospective study of AI triage for intracranial hemorrhage found no significant difference in report turnaround time with or without AI (about 147 to 150 minutes on average), suggesting that workflow context matters as much as the algorithm itself.
- Organ segmentation: Before radiation therapy, clinicians must outline every organ near a tumor so the radiation beam avoids them. This contouring work traditionally takes about 20 minutes per patient by hand. Deep learning-based auto-contouring reduces that time by a median of 71%, freeing clinicians for higher-value tasks.
- Report generation: Newer systems combine image analysis with language models to draft radiology reports. One approach generates a rough draft from the scan, then refines it by correcting errors and improving clinical accuracy. Another system called KARGEN merges visual features with a medical knowledge graph to produce detailed, clinically grounded reports. These tools are still being validated but represent a significant shift in how radiology documentation could work.
Opportunistic Screening
One of the most promising applications is something called opportunistic screening. Every year, millions of CT scans are performed for routine reasons: checking for kidney stones, evaluating abdominal pain, staging cancer. Those scans contain a wealth of information about bone density, muscle mass, liver fat, and arterial calcification that typically goes unreported because it wasn’t the reason for the scan.
AI can automatically measure these markers in the background. A routine abdominal CT could simultaneously flag early osteoporosis, detect fatty liver disease, quantify visceral fat linked to metabolic syndrome, or measure calcium buildup in arteries that signals cardiovascular risk. Cardiovascular disease, osteoporosis, metabolic syndrome, and sarcopenia (age-related muscle loss) account for the majority of deaths in aging populations, and catching any of these before symptoms appear could meaningfully improve outcomes. Early evidence suggests this kind of automated screening is not only clinically useful but actually cost-saving compared with ignoring the data.
Known Limitations and Risks
AI in radiology is only as reliable as the data it learned from, and that creates real blind spots. If a training dataset is dominated by images from younger patients, the system may struggle to recognize age-related conditions like diverticulosis in older adults. If the data skew toward one racial or ethnic group, accuracy drops for underrepresented populations. Even the specific scanners and imaging protocols used at the institutions that supplied training data can introduce bias, meaning an algorithm that performs well at one hospital may underperform at another using different equipment.
These aren’t theoretical concerns. Demographic imbalances in training data can lead to misdiagnoses or delayed diagnoses for the exact populations that already face healthcare disparities. The challenge is compounded by the fact that most commercial AI tools don’t disclose the demographic breakdown of their training sets, making it difficult for hospitals to assess how well a product will perform on their specific patient population.
Impact on Radiologists
The assumption has been that AI will reduce radiologist workload and burnout. The reality is more complicated. A large cross-sectional study of radiologists in China found that frequent AI use was actually associated with higher rates of burnout, with a dose-response pattern: the more regularly radiologists used AI, the stronger the association. This effect was most pronounced among radiologists with already high workloads and those with lower acceptance of the technology.
That finding likely reflects the current state of integration rather than an inherent flaw in the technology. When AI tools generate alerts that need to be reviewed, add steps to existing workflows, or produce false positives that require investigation, they can create more work rather than less. Radiologists who are skeptical of the tools may find the added friction especially draining. The study also found that high AI acceptance partially offset the burnout association, suggesting that training, trust, and thoughtful implementation matter as much as the algorithm’s accuracy.
AI is not replacing radiologists. It is changing what their daily work looks like, shifting time away from certain repetitive tasks and toward reviewing AI outputs, handling more complex cases, and integrating new types of automated findings into patient care. How well that transition goes depends heavily on how these tools are built, validated, and woven into clinical workflows.