AI is helping radiologists extract hidden information from routine medical images, revealing subtle biomarkers that could identify cancer and other diseases long before conventional diagnosis.
For many patients, pancreatic cancer produces no warning signs until it has spread, when treatment options are limited, and survival rates are among the lowest of any major cancer. But radiologists have long wondered whether the disease leaves clues earlier—before a tumor can be seen on a scan. Until recently, those signals were buried in the millions of pixels that make up a CT image. Now, advances in AI suggest they may have been visible all along.
Researchers at Mayo Clinic recently developed an AI model capable of detecting subtle tissue changes on routine abdominal CT scans months—and, in some cases, years—before pancreatic cancer is clinically diagnosed. At the same time, investigators at other institutions are using similar AI techniques to estimate future lung and breast cancer risk and uncover musculoskeletal abnormalities on routine chest radiographs that often go unrecognized.
Together, projects such as these point to a fundamental shift in medical imaging. Rather than simply identifying disease after it becomes visible, researchers are exploring whether AI can transform everyday imaging studies into predictive tools that identify patients who are at an elevated risk while intervention is still possible.
Hidden Patterns
The Mayo Clinic model, known as the Radiomics-based Early Detection Model (REDMOD), was designed to analyze routine abdominal CT examinations that originally appeared normal. Instead of searching for an obvious mass, the algorithm evaluates hundreds of quantitative imaging features describing tissue texture and architecture patterns too subtle for radiologists to perceive visually that may reflect the earliest biological changes associated with pancreatic cancer.
“The idea grew out of a clinical frustration,” says Ajit Goenka, MD, a professor of radiology and nuclear medicine specialist at Mayo Clinic. “We kept seeing patients diagnosed with pancreatic cancer who had CT scans months or years earlier that were read as normal. The cancer was there. It just had not declared itself visually.”
The question, he says, became whether apparently normal scans contained quantitative information that machine learning could detect, even when experienced radiologists could not. That question launched a research effort in 2019. An initial proofof- concept study published in Gastroenterology in 2022 demonstrated the feasibility of the approach. REDMOD, recently described in Gut, represents nearly six years of refinement involving automated segmentation, feature engineering, classifier development, and multi-institutional validation.
Unlike many AI models trained using a single institutional dataset, REDMOD was evaluated using CT examinations acquired at multiple hospitals on different scanners using varying imaging protocols. Researchers analyzed nearly 2,000 CT studies, including scans from patients who would later develop pancreatic cancer but whose examinations had originally been interpreted as normal.
The AI system correctly identified 73% of prediagnostic pancreatic cancers at a median of approximately 16 months before clinical diagnosis—nearly doubling the detection rate achieved by radiologists reviewing the same studies without AI assistance. Its advantage became even more pronounced with earlier examinations, Goenka says. On CT scans performed more than two years before diagnosis, REDMOD detected nearly three times as many future cancers as conventional image review alone, he says.
Perhaps equally important, Goenka says, the algorithm produced consistent results across serial examinations, suggesting that AI-generated risk assessments could eventually support longitudinal monitoring of patients considered at elevated risk. “The greatest barrier to saving lives from pancreatic cancer has been our inability to see the disease when it is still curable,” he says. “This AI can now identify the signature of cancer from a normal-appearing pancreas, and it can do so reliably over time and across diverse clinical settings.”
A Risk Signal, Not a Diagnosis
Despite the encouraging findings, Goenka is careful to distinguish between identifying elevated risk and diagnosing cancer. “It is a risk signal, not a diagnosis,” he says. Rather than confirming malignancy, REDMOD identifies tissue characteristics resembling those seen in patients who later developed pancreatic cancer. A positive result would indicate that additional evaluation is warranted, with physicians determining the appropriate next steps based on the patient’s clinical history and overall risk profile, Goenka says. Potential follow-up could include repeat CT imaging, endoscopic ultrasound, molecular PET imaging, tissue sampling, or continued surveillance, depending on the individual patient.
The technology is intended primarily for patients with newly diagnosed diabetes after age 50 who are identified as being at elevated risk through validated risk scores, Goenka says. Familial pancreatic cancer represents a different clinical population already monitored through established surveillance programs involving MRI and endoscopic ultrasound, he notes.
Researchers are now evaluating the technology prospectively through the Artificial Intelligence for Pancreatic Cancer Early Detection (AI-PACED) trial, which is examining how AI-guided risk assessment performs in real-world clinical practice. Investigators will study not only early detection rates, but also falsepositive findings, workflow integration, and patient outcomes.
If prospective studies confirm the retrospective results, the implications could extend well beyond pancreatic cancer, Goenka says. “The question for each organ is whether early carcinogenesis produces a quantitative texture signature that is detectable before a visible lesion forms,” he says. “That is an empirical question, not a theoretical one, and the answer will differ by disease biology.” While Goenka does not want to speculate beyond his group’s data, he adds that he believes “if a cancer has a desmoplastic or stromal reaction during its preclinical phase, there is reason to investigate.”
The Mayo researchers also are looking at analyses to determine whether the AI signal corresponds to the location where a pancreatic tumor eventually develops, comparisons between REDMOD and newer foundation AI models, and economic modeling to evaluate whether AI-assisted screening could be costeffective for health care systems. “We’re working on the clinical and institutional infrastructure needed to deliver this to the patients who need it,” Goenka says.
New Opportunities
While the Mayo Clinic team is focused on identifying one of the deadliest cancers before it becomes clinically apparent, researchers elsewhere are asking a broader question: How much untapped information is already hidden in routine medical images?
That question is driving several projects supported through HOPPR’s Catalyst Program. HOPPR was founded in 2019 to bring together experts in clinical radiology, AI development, and health care commercialization to advance the development of transparent and scalable AI for medical imaging. The program provides researchers access to foundation AI models, computing resources, and development tools that allow clinicians to adapt existing imaging AI to new clinical questions without building algorithms entirely from scratch. Its foundation models are trained on enormous numbers of medical images to learn general imaging characteristics. They can then be “fine-tuned” for specific applications, much as large language models can be adapted for specialized tasks.
Among the first of its Catalyst projects is underway at the University of Illinois Cancer Center in Chicago, where investigators are exploring whether chest CT scans and mammograms already obtained for screening can provide clues about a person’s future cancer risk. For Ameen Salahudeen, MD, PhD, assistant professor at the University of Illinois Cancer Center, the motivation comes directly from clinical experience. “As an oncologist, it’s very clear how much early detection is able to save lives,” he says. “It’s a different story when a tumor is detected via screening vs a cancer that is discovered when a patient has symptoms from metastatic disease.”
Current screening programs largely identify cancers that already exist. Salahudeen hopes AI can help identify patients most likely to develop cancer years before a lesion becomes detectable. “Our thesis is focused on how understanding long-term risk can help people make informed health and preventive decisions,” Salahudeen says.
Lung cancer, for example, remains the leading cause of cancer death worldwide, despite the availability of effective screening for eligible patients. Yet participation in annual low-dose CT screening remains disappointingly low, and many patients fail to return for subsequent examinations. “We think the biggest win could simply be helping patients prioritize their health and getting the recommended follow-up screenings,” Salahudeen says. Similarly, he adds, providing women with individualized breast cancer risk information derived from screening mammograms could encourage adherence to future screening recommendations while helping physicians identify patients who may benefit from closer surveillance.
Like Goenka, however, Salahudeen emphasizes that these AI systems are intended to support—not replace— clinical judgment. “There is no crystal ball,” he says. “We generally believe that with the results from labs, screenings, findings from the AI models, and other information, there are choices patients can make to reduce their risk.” Salahudeen envisions AI becoming another tool that physicians incorporate into shared decision-making discussions, much like cardiovascular risk calculators are used today.
Unlocking Routine Images
Cancer prediction is only one example of how researchers hope foundation models can expand the clinical value of existing imaging examinations. At The Catholic University of Korea, investigators are adapting a chest radiograph foundation model to detect thoracic spine abnormalities that frequently appear on routine chest X-rays but may receive little attention because the examination was ordered to evaluate cardiopulmonary disease.
“Chest X-rays contain much more information than cardiopulmonary findings alone,” says Joon-Yong Jung, MD, PhD. As a musculoskeletal radiologist, Jung says he frequently noticed spinal abnormalities that were visible on chest radiographs but often remained secondary to the primary clinical indication.
His team’s retrospective study uses chest radiographs validated against dedicated spine radiographs, CT, and MRI examinations to determine whether AI can reliably recognize thoracic spine abnormalities using relatively simple image preprocessing techniques. The project remains in its early stages and has been submitted for presentation at RSNA 2026, but Jung believes it illustrates a broader opportunity for what many researchers now call opportunistic screening, in which AI extracts clinically useful information from scans patients have already undergone.
Importantly, Jung stresses that the software would not establish a diagnosis. “A high-risk result would not be considered a diagnosis by itself,” he says. Instead, it would prompt radiologists and referring physicians to correlate the findings with the patient’s history and determine whether dedicated imaging or specialty referral is appropriate.
From Detection to Prediction
While the Mayo Clinic pancreatic cancer research has generated excitement, Paul Chang, MD, professor of radiology and vice chair of radiology informatics at the University of Chicago, sees the work as part of a broader evolution toward opportunistic screening. He envisions algorithms automatically quantifying coronary artery calcium, fatty liver disease, sarcopenia, and other biomarkers that would be impractical for radiologists to measure manually.
Chang draws a distinction, however, between AI applications that quantify findings already visible on imaging and newer algorithms that attempt to predict future disease risk from subtle imaging patterns. He describes the latter as an intriguing but very early area of research. “These studies are interesting, but we don’t yet know whether the models will perform reliably across different patient populations,” he says.
Even if predictive imaging algorithms ultimately prove accurate, Chang says significant practical and ethical questions remain. Beyond demonstrating that they improve outcomes, researchers will need to determine who pays for the technology and how clinicians should act on predictions of future disease. Informing a patient that AI estimates an elevated risk of pancreatic cancer years before symptoms appear could trigger anxiety, repeated imaging, unnecessary costs, and even unintended consequences involving insurance or employment. “Just because we can make a prediction doesn’t automatically mean it’s beneficial,” Chang says. “We have to prove that these tools improve patient care, not simply that they generate statistically significant predictions.”
Curtis P. Langlotz, MD, PhD, professor of radiology and director of the Center for Artificial Intelligence in Medicine and Imaging at Stanford University, agrees that predicting future disease risk represents the next frontier for imaging AI. “AI has become very good at detecting abnormalities already visible on imaging,” Langlotz says. “Predicting future disease risk from routine imaging is one of the next frontiers. Accurate predictions will require multimodal data— not just imaging data—in most cases.”
Langlotz says studies such as the Mayo Clinic’s suggest AI may be uncovering biologic signals that were previously invisible to radiologists. “This surpasses what we previously thought were the limits of medical imaging,” Langlotz says. “It will be interesting to see which diseases have image signatures that have not been detected previously by human eyes.”
Still, Langlotz says, translating these advances into everyday practice will require more than sophisticated algorithms. Although retrospective studies are encouraging, Langlotz says prospective trials must show that predictive AI improves outcomes. Equally important, Langlotz says, health systems need workflows defining who receives risk information, how it is communicated, and what actions should follow.
Beth W. Orenstein of Northampton, Pennsylvania, is a freelance medical writer and regular contributor to Radiology Today.