Why misdiagnosis carries serious consequences: Autoimmune pancreatitis (AIP) is a rare inflammatory condition of the pancreas that can look almost identical to pancreatic ductal adenocarcinoma (PDA) on medical imaging. Patients with AIP can be cured with steroids, while pancreatic cancer typically requires surgery or chemotherapy. Misdiagnosing AIP as cancer can lead to unnecessary, harmful surgery, while misdiagnosing cancer as AIP delays critical treatment.
Why current imaging falls short: Both AIP and pancreatic cancer can present as a mass in the pancreas, may cause similar symptoms including jaundice and abdominal pain, and can look similar on conventional CT and MRI. Serum markers like IgG4 help in some cases, but they are not definitive, and tissue biopsy of the pancreas carries procedural risk. A reliable, non-invasive imaging-based test to distinguish the two conditions would be enormously valuable clinically.
PET/CT imaging as a multi-modal approach: Combined PET/CT imaging provides two complementary information streams. CT shows anatomical structure and tissue density, while PET using the FDG glucose tracer shows metabolic activity. Cancer cells tend to be highly metabolically active on PET, while AIP shows a more diffuse inflammatory pattern. Combining both modalities and extracting quantitative texture features may reveal discriminating patterns invisible to the human eye.
Machine learning as a discriminator: Support vector machines (SVMs) are machine learning models well-suited to classification problems where many features need to be combined into a single decision boundary. By training an SVM on PET/CT texture features from patients with confirmed diagnoses, the model can learn the multi-dimensional imaging signature that distinguishes AIP from PDA.
Patient cohort: The study included 45 patients with confirmed AIP and 66 patients with confirmed PDA, all of whom had undergone FDG-PET/CT imaging. Diagnoses were confirmed by pathology or by clinical criteria including response to steroid treatment for AIP cases. This real-world clinical cohort reflects the diagnostic challenge encountered in practice.
Feature extraction process: Radiologists manually delineated the pancreatic lesion regions of interest on both the CT and PET images for each patient. From these regions, 418 texture features were extracted using standardized radiomics analysis tools, capturing patterns related to signal intensity distribution, local spatial relationships, and higher-order statistical properties of the image texture.
Feature selection strategy: Reducing 418 features to a manageable and non-redundant set was essential to avoid overfitting. The Fisher criterion was first applied to rank features by their ability to separate the two classes, followed by Sequential Forward Floating Selection (SFFS) combined with SVM cross-validation to identify the optimal subset. This iterative approach selected 8 final features from the original 418.
SVM training and validation: The SVM classifier was trained on the selected 8 features using leave-one-out cross-validation, where each patient was held out once as a test case while the model was trained on all other patients. This provides an unbiased estimate of how the model would perform on new patients not seen during training.
Classification performance: The final SVM model using 8 selected PET/CT texture features achieved an area under the ROC curve (AUC) of 0.9218 and an overall accuracy of 89.28% for distinguishing AIP from PDA. These are strong results for a task that frequently challenges experienced clinicians using conventional imaging interpretation.
PET features outperformed CT features: When the discriminative power of individual features was analyzed, PET-derived texture features consistently ranked higher than CT-derived features. This finding suggests that metabolic activity patterns captured by FDG-PET carry more diagnostic information for AIP versus PDA differentiation than anatomical tissue density patterns from CT alone.
The 8 selected features: The final feature set included texture metrics from both CT and PET modalities, capturing properties such as intensity uniformity, spatial heterogeneity, and run-length patterns. These features collectively encode both the metabolic heterogeneity visible on PET and the structural properties of the lesion on CT.
Performance relative to clinical judgment: The model's AUC of 0.9218 compares favorably with reported diagnostic accuracy of conventional imaging interpretation and serum IgG4 testing in this differential diagnosis scenario. This suggests the model captures diagnostically relevant information that may be difficult to fully leverage through subjective visual assessment.
High stakes for patients: Pancreatic surgery (the Whipple procedure or distal pancreatectomy) carries significant morbidity and mortality risks. For patients with AIP who are mistakenly operated on as cancer cases, these risks are entirely avoidable. A reliable pre-operative test to confidently diagnose AIP could prevent hundreds of unnecessary surgeries annually in high-volume medical centers.
Complementing existing diagnostic criteria: The current standard for diagnosing AIP uses the ICDC (International Consensus Diagnostic Criteria), which incorporates imaging, serology, histology, and steroid response. The PET/CT texture feature approach provides an additional quantitative imaging criterion that could strengthen diagnostic confidence in ambiguous cases where the ICDC criteria are not clearly met.
Potential for reduced use of risky biopsies: Confirming AIP sometimes requires endoscopic ultrasound-guided biopsy, which carries procedural risks. A highly accurate non-invasive PET/CT-based classifier could provide sufficient diagnostic confidence to avoid biopsy in a subset of cases, reducing procedural risk and cost.
The power of quantitative imaging features: This study demonstrates that numbers extracted from medical images, specifically texture statistics that quantify patterns invisible to casual visual inspection, can encode clinically important diagnostic information. The radiomics approach treats imaging data as a quantitative measurement rather than a picture to be interpreted subjectively.
Reproducibility and standardization: One advantage of a model based on standardized texture features and an SVM classifier is that it is highly reproducible. Once trained and validated, the same model can be applied to new cases with consistent results, independent of the experience level of the clinician reading the images.
Future directions: The current model requires manual lesion segmentation, which is time-consuming. Future work could incorporate automated segmentation to create a fully automated diagnostic pipeline. External validation in independent patient cohorts from other institutions is also needed before this approach could be integrated into clinical practice guidelines.