Abstract
The radiomics models based on multiparametric MRI have the potential ability to classify PDAC and MFCP lesions.
Introduction
Pancreatic ductal adenocarcinoma (PDAC) is among the most lethal cancers, yet its diagnosis is frequently complicated by mass-forming chronic pancreatitis (MFCP). MFCP is a benign inflammatory condition that often mimics the appearance of malignant tumors on standard CT or MRI scans. This similarity leads to a significant clinical dilemma: approximately 5% to 11% of patients undergoing major pancreatic surgery are post-operatively found to have had benign inflammation rather than cancer. Conversely, misdiagnosing malignancy as inflammation delays life-saving treatment. Because pancreatic biopsies are invasive and prone to false negatives, there is a critical need for a non-invasive method to differentiate these conditions.
Methods
Researchers developed a machine-learning framework utilizing radiomics—the extraction of high-dimensional data from medical images. The study followed a rigorous workflow:
- Imaging: Multiparametric MRI was used, including T1-weighted (T1WI), T2-weighted (T2WI), and dynamic contrast-enhanced (DCE) scans (Arterial and Portal phases).
- Segmentation and Extraction: Radiologists manually outlined the "Region of Interest" (ROI) on the largest cross-section of the lesion. Using IBEX software, 410 features were extracted per sequence, including intensity histograms and texture matrices.
- Preprocessing and Selection: Images were normalized, and a Laplacian-of-Gaussian (LoG) filter was applied to highlight fine-grained textures. The LASSO (Least Absolute Shrinkage and Selection Operator) algorithm was employed to select the most predictive features.
- Classification: A Support Vector Machine (SVM) with a Gaussian kernel was trained to classify the lesions.
The study utilized a retrospective dataset of 119 patients from two independent institutions in China, split into a training cohort (n=64) and an external testing cohort (n=55).
Key Results
The radiomics models significantly outperformed both clinical biomarkers (like CA19-9) and experienced radiologists. The model based on the portal-venous phase of the MRI achieved the highest performance, suggesting that the contrast wash-through during this window provides the most discriminative texture for identifying malignancy.
| Model | Training AUC | Validation AUC | Sensitivity (Val) | Specificity (Val) |
|---|---|---|---|---|
| T1WI Model | 0.893 | 0.882 | - | - |
| T2WI Model | 0.911 | 0.902 | - | - |
| Arterial Phase | 0.958 | 0.920 | - | - |
| Portal Phase | 0.997 | 0.962 | 97.8% | 90.0% |
| Clinical Model | 0.516 | 0.649 | - | - |
Discussion and Conclusion
The core finding is that while PDAC and MFCP may appear as similar "gray blobs" to the human eye, their underlying micro-structures differ in ways that radiomics can quantify. Traditional clinical metrics, such as lesion size and duct diameter, performed poorly (AUC ~0.52–0.65), highlighting the limitations of standard diagnostic criteria.
While the study is limited by its retrospective nature and a relatively small sample of MFCP cases, it demonstrates that SVM models can maintain high sensitivity while significantly reducing false positives compared to human interpretation. This radiomics approach offers a promising non-invasive path toward reducing unnecessary surgeries and ensuring timely treatment for pancreatic cancer patients.
Summary updated