MRI heterogeneity maps predict pancreatic cancer nerve invasion

A machine-learning framework that reads spatial patterns across three MRI sequences predicted perineural invasion in pancreatic ductal adenocarcinoma with strong discrimination in two independent hospital cohorts, according to a multicenter retrospective study. The combined imaging-and-clinical model achieved an area under the receiver operating characteristic curve, or AUC, of 0.860 in one external validation cohort and 0.869 in the other. Its internal test AUC was 0.907. These results suggest that the microscopic spread of cancer into or around nerves may leave quantifiable, spatially fragmented signatures that conventional scans do not reveal directly.

Perineural invasion is the infiltration of tumor cells into the epineurium, perineurium or endoneurium surrounding peripheral nerves. It was identified in 137 of the 205 surgically treated patients included in the study, while 68 had no perineural invasion. The condition is associated in prior research with local recurrence, postoperative neuropathic pain and impaired survival, but definitive classification currently depends on examination of the surgically removed specimen. The new study asked whether information available before surgery could estimate that microscopic finding noninvasively, potentially giving surgeons and oncologists additional risk-stratification information while decisions are still being made.

To do so, the researchers analyzed axial T2-weighted MRI, diffusion-weighted MRI and contrast-enhanced T1-weighted MRI obtained within one month before surgery. Two radiologists manually delineated each tumor on every sequence, excluding necrosis, cystic components and major blood vessels. Their agreement was assessed with an intraclass correlation coefficient, and only features with a coefficient above 0.80 were retained. Images were resampled to 1 × 1 × 1 mm isotropic voxels and processed with N4 bias-field correction to reduce intensity inhomogeneity. The resulting volumes were then used to characterize intratumoral heterogeneity, meaning the spatial diversity of different tissue patterns within a single tumor.

The team extended a validated spatial heterogeneity framework in two dimensions and three dimensions. For the two-dimensional analysis, the axial slice with the largest tumor area was selected and 104 radiomic features—shape, first-order intensity and texture measurements—were extracted from each pixel. K-means clustering partitioned the tumor into K groups of similar imaging characteristics, with robustness examined across K values from 3 to 9. A score based on the number and size of spatially connected clusters measured how fragmented those labels were across the tumor. Higher scores indicated greater spatial heterogeneity. The three-dimensional procedure applied the same concept to voxel neighborhoods, producing a core heterogeneity score plus 15 complementary measures of volume, entropy, intensity range, boundary complexity and fractal structure.

A two-stage selection process reduced the imaging variables to nine features. Five came from T2-weighted MRI, one from diffusion-weighted MRI and three from contrast-enhanced T1-weighted MRI. LASSO logistic regression first retained variables with nonzero coefficients within each modality, and random-forest-based recursive feature elimination then condensed the combined set. The investigators also tested 11 classification algorithms: logistic regression, random forest, support vector machine, extreme gradient boosting, light gradient boosting machine, k-nearest neighbors, decision tree, gradient boosting decision tree, adaptive boosting, naive Bayes and linear discriminant analysis. Hyperparameters were optimized through five-fold stratified cross-validation, with balancing strategies used where needed to address the unequal numbers of patients with and without invasion.

Multisequence models generally performed better than models based on a single MRI sequence. T2-weighted imaging produced the strongest single-modality results, although its best internal-test AUC was only 0.723; the corresponding descriptive maxima in the two external cohorts were 0.789 and 0.788. By contrast, the best combined-sequence random forest reached AUCs of 0.881 in the internal test set, 0.860 in the first external cohort and 0.850 in the second. The training cohort was split from patients at two centers, with 100 patients assigned to training and 43 to internal testing, while separate surgical cohorts of 28 and 34 patients from two additional centers served as external validation groups. That four-center arrangement was important because it tested whether performance persisted across hospitals, scanner manufacturers and acquisition protocols.

The study then compared a clinical model, an imaging model and a combined model. The clinical model used tumor location and preoperative clinical T-stage, the only two baseline variables independently associated with perineural invasion in logistic-regression analysis. Tumors in the pancreatic body or tail had lower adjusted odds of invasion than tumors in the pancreatic head, with an adjusted odds ratio of 0.36 and a 95% confidence interval of 0.20 to 0.67. Advanced clinical stage, defined as T3 or T4 rather than T1 or T2, was associated with higher odds, with an adjusted odds ratio of 2.26 and a 95% confidence interval of 1.18 to 4.30. Yet the clinical model alone was unstable outside the training hospitals, producing AUCs of 0.758 in internal testing, 0.342 in the first external cohort and 0.700 in the second.

By comparison, the nine-feature imaging model recorded AUCs of 0.881, 0.860 and 0.850 in the internal and two external test sets. Adding tumor location and clinical stage raised the internal AUC to 0.907 and the second external AUC to 0.869, while the first external AUC remained 0.860. Decision-curve analysis indicated greater modeled clinical net benefit from the imaging and combined approaches than from the clinical model across most threshold probabilities. Calibration results varied among cohorts, with greater uncertainty in the small external groups. The findings therefore support the predictive contribution of MRI heterogeneity beyond the two clinical variables, but the unusually low clinical-model AUC may also reflect sampling variation or differences in case mix and stage documentation; the study could not establish the cause.

To make the combined model more transparent, the researchers used SHAP values, which quantify how individual variables influence a prediction. The three largest average contributions came from T2-weighted features: the three-dimensional maximum subregion proportion, a two-dimensional heterogeneity score based on six clusters, and a three-dimensional intensity range metric. Clinical T-stage and a T1-weighted intensity-range feature followed, while tumor location made the smallest contribution. The direction of influence was not uniform or monotonic. For example, a larger connected tumor subregion generally lowered the model’s contribution to predicting invasion. The authors emphasized that these are model-specific imaging associations: they do not establish a particular clonal mechanism or prove that a specific tissue pattern causes nerves to become involved.

Several limitations constrain how far the results can be generalized. The cohort was retrospective and included only patients who underwent resection with pathology available for confirming invasion, so the method was not tested in unresectable disease, borderline resectable disease or other populations that may be more relevant to preoperative neoadjuvant treatment. Manual tumor delineation requires radiologist time, and automated segmentation was not evaluated. MRI scanners and protocols differed among the four centers; resampling, bias-field correction and standardization addressed some differences, but no feature-level batch-effect correction such as ComBat was applied. Most importantly, the external cohorts contained only 28 and 34 patients, making their performance estimates imprecise. Larger prospective studies would be needed to test fixed thresholds, calibration and clinical utility before such predictions could guide surgery or treatment. Within those bounds, the study shows that a map of spatial variation across multiple MRI sequences can provide a promising preoperative surrogate for a microscopic feature of pancreatic cancer.

Subject of Research: Preoperative prediction of pancreatic ductal adenocarcinoma perineural invasion using multisequence MRI heterogeneity and machine learning

Article Title: Machine learning models integrating 2D and 3D intratumoral heterogeneity features from multi-sequence MRI: accurate prediction of perineural invasion in pancreatic ductal adenocarcinoma

Article References: Ouyang, Z., Song, Z., Li, Z., Cui, S., Liu, X., Shen, W., Zhu, F., Liao, C., Hu, W., He, S., Lu, Y., & Ye, Q. (2026). Machine learning models integrating 2D and 3D intratumoral heterogeneity features from multi-sequence MRI: accurate prediction of perineural invasion in pancreatic ductal adenocarcinoma. Frontiers in Artificial Intelligence, 9, Article 1975752. https://doi.org/10.3389/frai.2026.1975752

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1975752

Keywords: pancreatic cancer, perineural invasion, magnetic resonance imaging, intratumoral heterogeneity, machine learning, radiomics, external validation, clinical imaging, Machine, learning, models, integrating