Gaussian Meta-Space Augmentation for Stacking Ensembles in Multimodal IPMN Risk Stratification

Pancreatic cancer is a highly lethal malignancy, and identifying malignant transformation risks in intraductal papillary mucinous neoplasms (IPMNs) is crucial for early intervention. Traditional risk assessment often requires invasive biopsies, while non-invasive imaging analysis using magnetic resonance imaging (MRI) faces challenges due to subtle patterns missed by human visual assessment. Although quantitative methods like radiomics and deep learning offer promise, effectively fusing information across different MRI sequences, anatomical regions, and feature representations remains a significant technical challenge without causing overfitting in downstream classifiers.

The authors evaluate a multi-center dataset of 678 unique patients with paired T1-weighted and T2-weighted MRI scans. The proposed pipeline treats IPMN risk stratification as a stacked generalization task. First, base classifiers are trained on individual streams, which combine feature representations (engineered radiomics versus convolutional neural network descriptors), anatomical regions (whole organ versus head, body, and tail sub-regions), and MRI sequences. These base models produce out-of-fold class probabilities.

Next, the authors introduce calibrated Upstream Probabilistic Meta-Imputation (cUPMI). This technique fits a class-conditional Gaussian distribution with a shared pooled covariance to the log-probability meta-features and injects synthetic samples upstream of the level-1 stacking classifier to smooth its decision surface. The study investigates various level-1 combiners, including L2-regularized logistic regression, random forests, and XGBoost, evaluating binary and ordinal (three-class: no-risk, low-risk, high-risk) risk prediction.

Evaluating across multiple prediction paradigms, the authors find that single deep-learning architectures, such as a 3D DenseNet-121, underperform compared to lighter 2.5D ResNet models and radiomics-based stacking ensembles. The highest overall performance is achieved by a fold-locked random forest stack over fused deep-learning and radiomics streams, which reaches a quadratic-weighted Cohen's kappa (QWK) of 0.595 and a binary AUC of 0.839.

Regarding the proposed cUPMI regularizer, its benefits scale with the capacity of the combiner. It provides negligible advantages for already-regularized L2-logistic stacks but consistently regularizes higher-capacity tree combiners. For instance, it yields consistent improvements in binary AUC for random forest and XGBoost models, and offers its cleanest ordinal benefit for XGBoost in an 8-stream radiomics task.