Pancreatic ductal adenocarcinoma is notoriously difficult to detect early. By the time most patients receive a diagnosis, the cancer has already spread and surgical cure is no longer possible. Earlier detection on routine CT scans, before tumors become visually obvious, could dramatically improve survival.
Radiomics-based machine learning models can analyze subtle patterns in CT images that are invisible to the human eye. Researchers at Mayo Clinic previously developed a support vector machine (SVM) model that could identify pre-diagnostic pancreatic CT scans—those taken before cancer became apparent. This study tests how reliably that model performs across the real-world variations that occur in clinical imaging.
The model was originally trained on CT scans of 45 patients who were later diagnosed with pancreatic cancer and 83 patients with healthy pancreases. To test its robustness, researchers created 18 additional modified versions of the test dataset by systematically altering images in ways that happen naturally in clinical practice.
The alterations covered five categories: increasing image noise (at 1x, 2x, 3x, and 5x levels), rotating images by 45 and 90 degrees, resampling voxel sizes in different ways, changing gray-level bit depth, and deliberately eroding or dilating the pancreas segmentation boundary by 3 to 8 pixels. Each perturbation was applied separately so researchers could isolate which changes affected model performance.
In the original unmodified test set, the model correctly classified 43 out of 45 pre-diagnostic cancer scans and 75 out of 83 normal scans, achieving 92.2% accuracy and an impressive AUC of 0.98. This baseline performance held up well under most perturbation scenarios.
Image rotation, voxel resampling, and moderate gray-level changes did not meaningfully affect performance. Similarly, eroding or dilating the pancreas boundary by up to 4 pixels did not hurt accuracy. The only notable decline was at extreme noise levels—when noise was tripled, sensitivity dropped to 80%, suggesting very poor image quality may challenge the model.
Radiomics involves extracting hundreds of quantitative features from medical images—things like texture, shape, and intensity patterns. Two features tied to CT slice thickness were removed during development to improve stability. The remaining 32 features formed the basis of the support vector machine model.
For a radiomics model to be useful in clinical care, it must produce consistent results regardless of which CT scanner was used, how the patient was positioned, or exactly how the radiologist drew the tumor boundary. This study's comprehensive perturbation testing directly addresses that requirement, providing evidence the model is ready for deployment across diverse clinical environments.
This validation work is a critical step before any AI diagnostic tool can be trusted in clinical settings. By showing the model holds up under conditions that vary from scan to scan and institution to institution, the research team demonstrates it is not a fragile lab-only tool but a robust detection system.
If deployed, such a model could scan existing pre-diagnostic CTs of high-risk patients—such as those with new-onset diabetes or unexplained weight loss—and flag pancreases showing subtle radiomics changes, triggering earlier follow-up. Catching pancreatic cancer even months earlier could make the difference between resectable and unresectable disease.
The Mayo Clinic SVM model demonstrated exceptional stability across nearly all tested variations in image acquisition and radiomics workflow. Only extreme noise conditions caused a meaningful drop, and such noise levels are rarely encountered in modern clinical CT scanners.
These results provide strong evidence that the model is suitable for prospective clinical validation studies. The next step would be testing the model in real-time clinical workflows across multiple institutions to confirm it can reliably identify patients who need earlier follow-up for pancreatic cancer.