Autoimmune pancreatitis (AIP) is a rare inflammatory condition where the immune system attacks the pancreas. When it appears as a focal mass, it looks nearly identical to pancreatic cancer on CT scans, creating a critical diagnostic problem.
If autoimmune pancreatitis is misdiagnosed as cancer, patients may undergo unnecessary and high-risk surgery. The Whipple procedure carries significant morbidity and mortality. Conversely, a cancer misdiagnosed as pancreatitis may be denied timely treatment.
AIP responds well to steroid therapy, sparing patients from surgery entirely. But distinguishing the two conditions by imaging alone is challenging even for expert radiologists, particularly when lesions are small.
Researchers collected CT scans from 96 patients — 45 with confirmed focal autoimmune pancreatitis and 51 with confirmed pancreatic ductal adenocarcinoma — from two institutions, all diagnosed by pathology, biopsy, or established AIP diagnostic criteria.
Each patient underwent multi-phase CT: unenhanced, arterial phase, and venous phase. From each phase, 1,160 radiomics features were automatically extracted from the pancreatic lesion region. Separate models were built per phase and the best-performing combination identified.
The radiomics model was directly compared to the diagnostic accuracy of two experienced radiologists who reviewed the same scans independently, allowing direct quantification of how much AI improved on expert human assessment.
The optimal radiomics model achieved sensitivity of 93.3%, specificity of 96.1%, and overall accuracy of 94.8% — exceptional results for a clinically challenging diagnostic problem that often stumps expert radiologists.
The two radiologists achieved sensitivities of 57.8% and 73.3%, specificities of 88.2% and 90.2%, and accuracies of 75.0% and 81.3% respectively. The AI outperformed both on all metrics.
The improvement in sensitivity is especially important clinically: misclassifying AIP as cancer leads to unnecessary surgery. The AI's ability to correctly identify more AIP cases means fewer patients would undergo a futile high-risk operation.
Pancreaticoduodenectomy is one of the most complex abdominal operations, with complication rates of 40-60% even at experienced centers and mortality rates of 1-3%. A patient with AIP who undergoes this surgery faces these risks for no benefit.
A radiomics tool achieving 94.8% accuracy could substantially reduce the number of AIP patients who are misdiagnosed and unnecessarily operated on. Since AIP responds dramatically to steroids with lesions often resolving within weeks, accurate diagnosis has immediate therapeutic consequences.
The multi-phase approach — combining unenhanced, arterial, and venous CT phase information — appeared important for high accuracy, as different phases highlight different aspects of tissue biology.
This study demonstrates that radiomics applied to standard multi-phase CT can solve one of the most challenging differential diagnoses in pancreatic medicine with accuracy substantially exceeding expert radiologists.
The approach uses standard CT scans already obtained as part of routine clinical workup, requiring no additional procedures. Clinical implementation is straightforward once the model is validated in larger prospective cohorts.
Next steps include prospective multi-institution validation, integration into radiology reporting workflows as a decision-support tool, and examination of performance in edge cases.