![\"CT](\"https:\/\/rtmedical.com.br\/wp-content\/uploads\/2026\/09\/tomografo-tc-abdome-rotina.jpg\")
Before arguing about the percentage, it is worth naming the category, because the category is the story. This is not diagnostic imaging in the conventional sense. It is opportunistic, predictive imaging: the CT was ordered for abdominal pain, for trauma, to follow a renal nodule, to stage something else entirely, and the algorithm arrives afterwards, on an archived study, to return a probability of future disease. The target is not a lesion. It is the biology that precedes the lesion. Subtle Signals<\/em>, the Radiology Today<\/em> feature that assembled these threads, pairs the pancreas work with parallel efforts in lung, breast and thoracic spine.<\/p>\n The programme dates to 2019, with a proof of concept in Gastroenterology<\/em> in 2022. The validated version represents close to six years of work on automated segmentation, feature engineering and classifier design. The cohort pairs 219 patients whose prediagnostic CT was acquired between 3 and 36 months ahead of a ductal adenocarcinoma diagnosis with 1,243 controls who had a normal pancreas and stayed cancer-free for at least three years, close to 2,000 studies in total. Training took 969 cases, testing 493, at roughly seven controls per case, and 71% of the prediagnostic CTs in the test set came from outside institutions, on other vendors and other protocols.<\/p>\n Segmentation runs through a three-dimensional nnU-Net. Of 968 candidate radiomic features, 40 survived selection, and roughly 90% of those derive from filtered rather than raw images. The final call comes from an ensemble of logistic regression, random forest and XGBoost. At the published operating point the model pairs 73% sensitivity with 81% specificity and an area under the curve of 0.82 using filtered features, against 0.74 without them. An independent external validation returned 87.5% specificity, and test-retest agreement across serial examinations of the same patient sat between 90% and 92%.<\/p>\n Two slices of the result deserve attention. On scans acquired more than two years before diagnosis, the model surfaced almost three times as many future cancers as conventional review. And the longitudinal stability suggests the score could be tracked over time rather than read once. Ajit Goenka, the Mayo radiologist and nuclear medicine specialist who leads the work, told Radiology Today<\/em> that the project grew out of a clinical frustration: patients arriving with pancreatic cancer whose earlier scans had been called normal. The disease was present, he said. It simply had not declared itself visually.<\/p>\n Radiomics turns an image into a table of numbers. Once the organ is outlined, the software derives intensity histograms, shape metrics and, above all, higher-order statistics describing how voxels sit relative to one another: coarseness, homogeneity, entropy, run lengths of similar grey values. Wavelet and Laplacian-of-Gaussian filters expose texture scales the eye cannot separate. None of this creates new information; it describes what the acquisition already recorded. The working hypothesis is that as pancreatic stroma begins to reorganise during the preclinical phase, the signature shows up in those statistics before it becomes a perceptible contour.<\/p>\n Goenka is explicit about the boundary. This is a risk signal, not a diagnosis. A positive result says that a pancreas resembles those of patients who later fell ill, and what follows depends on clinical history and risk profile: a repeat CT, endoscopic ultrasound, molecular PET, tissue sampling, or simply continued surveillance. The intended population is deliberately narrow, centred on new-onset diabetes after the age of 50 combined with a validated risk score. Familial pancreatic cancer is a different problem, already served by MRI and endoscopic ultrasound surveillance programmes.<\/p>\n The thesis is not confined to the pancreas. We have covered the AI that outperformed radiologists on early pancreatic cancer<\/a>, the model that flags breast cancer risk a decade ahead<\/a> from screening mammograms, and a radiomics model that grades hand osteoarthritis on plain radiographs<\/a>. The same family includes work showing that incidental findings on lung CT signal cancer risk<\/a>. The common thread is extracting more from an examination than the question that prompted it.<\/p>\n Other groups in the feature follow the same logic. At the University of Illinois Cancer Center in Chicago, Ameen Salahudeen is testing whether chest CT and mammography already performed carry usable signals about future cancer risk, on the practical bet that an individualised estimate improves adherence to follow-up. At The Catholic University of Korea, Joon-Yong Jung is adapting a chest radiograph foundation model to recognise thoracic spine abnormalities that appear on the film but rarely get attention, because the study was ordered for cardiopulmonary disease; the work has been submitted to RSNA 2026. Paul Chang of the University of Chicago notes that a good deal of opportunistic quantification is feasible today for coronary calcium, hepatic steatosis and sarcopenia, measurements nobody would perform manually at scale.<\/p>\n In Brazil the economic case is unusually clean, because the scan has already been paid for. Where there is no budget for a population screening programme, re-reading an abdominal CT that already sits in the archive costs compute, not acquisition. The obstacle sits elsewhere. It demands genuine integration between PACS, RIS and the model, so the output reaches the right person instead of dying in a parallel report; it demands a referral pathway agreed before the algorithm is switched on; and it demands an answer to the uncomfortable question of who owns a risk flag nobody requested. The radiologist who signed a normal abdominal CT was not asked about the pancreas, and if the score turns up later, somebody has to communicate it, follow it and pay for the work-up.<\/p>\n Three caveats survive any press release. The first is design. A retrospective case-control study with seven controls per case does not reproduce real prevalence, and the reported positive predictive value near 36% collapses when the same sensitivity and specificity meet an unselected population. At 81% specificity, roughly one in five healthy pancreases gets flagged, and every flag costs imaging, time and anxiety. The second is generalisation: radiomic features are notoriously sensitive to vendor, slice thickness, reconstruction kernel and contrast phase. Drawing 90% of features from filtered images and testing on 71% external scans helps, but it does not replace validation across a genuinely heterogeneous installed base.<\/p>\n The third is lead-time bias. Detecting something 16 months earlier only matters if what happens in those 16 months changes the outcome; otherwise the patient simply spends longer knowing. That is why the Mayo group moved the model into AI-PACED, a prospective feasibility study of automated risk stratification, serial AI-augmented imaging and biobanking in new-onset diabetes, with time to diagnosis as the primary endpoint. The design keeps a hard firewall: clinical reports come from radiologists blinded to the study and enter the record normally, while the AI analysis runs on de-identified data and never touches care decisions.<\/p>\n Stanford’s Curtis Langlotz frames the state of play plainly. AI has become strong at detecting abnormalities already visible, and predicting future disease is the next frontier, one that will usually require multimodal data rather than images alone. Chang is blunter about the risk of getting ahead of the evidence, arguing that being able to generate a prediction does not by itself make the prediction useful, and that these tools will have to demonstrate better patient care rather than statistical significance. Until that evidence exists, REDMOD is best read for what it is: the strongest demonstration so far that clinically useful information is sitting inside examinations we already call normal.<\/p>\nInside the numbers: 1,462 patients, 968 features, a 7:1 control ratio<\/h2>\n
What radiomics actually computes, and why a score is not a diagnosis<\/h2>\n
What it would change in the reading room<\/h2>\n
Where this could break: study design, kernels and lead time<\/h2>\n