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

The paper introduces cUPMI (calibrated Upstream Probabilistic Meta-Imputation), a class-conditional Gaussian augmentation method for regularizing level-1 stacking ensembles in multimodal IPMN (intraductal papillary mucinous neoplasm) risk stratification. The study uses a multi-center cohort of 678 patients from the Cyst-X collection, with paired T1W/T2W MRI sequences and whole-organ (n=629) and sub-regional (n=616) pancreas segmentations. The pipeline stacks base classifiers—radiomics random forests and 2.5D ImageNet-pretrained ResNet-18 deep learning streams—across 16 streams defined by representation (radiomics or CNN), sequence (T1W/T2W), and region (whole, head, body, tail). The meta-features are per-class log-probabilities concatenated across streams, and cUPMI fits one class-conditional Gaussian per class with a shared pooled covariance (ridge λ=10−4) in log-probability space, then injects synthetic samples (ratio ρ selected by inner cross-validation) into the training fold of the level-1 combiner. Three combiners are evaluated: L2-regularized logistic regression, random forest (200 trees, max depth 4), and XGBoost (200 shallow trees, max depth 3, learning rate 0.05). Binary classification uses AUC; ordinal 3-class (no<low<high) uses quadratic-weighted Cohen's κ (QWK) with macro one-vs-rest AUC.

Key results: cUPMI provides no benefit to the already-regularized L2 logistic stack in binary tasks (∆+0.0005 AUC) but consistently improves higher-capacity tree combiners in binary and radiomics-only settings—RF gains +0.015 and XGBoost +0.024 binary AUC, positive in all seeds. In ordinal classification, cUPMI's clearest benefit is for XGBoost on an 8-stream radiomics task (+0.022 QWK, positive in all seeds), while RF (+0.004) and LR (+0.007) show minimal gains. In the fused DL+radiomics setting, cUPMI fails to improve RF (∆QWK −0.003) and slightly hurts LR (−0.007), with XGBoost retaining a negligible +0.005 gain. The strongest overall model is a fold-locked RF stack over fused DL+radiomics without cUPMI, achieving QWK 0.595 (95% CI [0.54, 0.64]) and binary AUC 0.839, surpassing radiomics (QWK 0.514, AUC 0.813), 2.5D ResNet-18 (QWK 0.497, AUC 0.784), and 3D DenseNet-121 (QWK 0.452, AUC 0.774) baselines. Errors concentrate on adjacent grades with near-zero two-step confusions (0.02/0.04). The paper concludes that cUPMI regularizes higher-capacity tree combiners, and that careful fusion of whole-organ and sub-regional radiomics with deep-learning streams outperforms heavier single architectures on small ordinal cohorts. External validation and full multiclass calibration are noted as future work.

Improvements to AI systems:

1. Add class-conditional Gaussian augmentation with pooled covariance to tree-based ensemble meta-learners (random forests, XGBoost) in stacking pipelines. This injects synthetic per-class log-probability samples during training, which regularizes high-capacity combiners and yields +0.015 to +0.024 AUC gains in binary tasks, with consistent positive results across all random seeds.

2. Use per-class log-probabilities as meta-features instead of raw predictions or hard labels when stacking heterogeneous base models (radiomics + deep learning). This preserves calibration information and enables the Gaussian augmentation to operate in a well-behaved, bounded space.

3. Apply cUPMI selectively—only for tree-based combiners, not for already-regularized linear models (e.g., L2 logistic regression). The method provides negligible benefit (+0.0005 AUC) for linear stacks, so computational cost can be avoided there.

4. For ordinal classification (e.g., no<low<high), use quadratic-weighted Cohen’s κ as the optimization target and evaluate cUPMI on radiomics-only streams—the clearest gain is +0.022 QWK for XGBoost on an 8-stream radiomics task, positive across all seeds. Avoid applying cUPMI to fused deep-learning+radiomics ordinal stacks, where it can slightly hurt (RF −0.003, LR −0.007).

5. Fuse whole-organ and sub-regional (head/body/tail) radiomics with 2.5D ImageNet-pretrained ResNet-18 streams using a fold-locked random forest combiner (without cUPMI) for the strongest overall performance: QWK 0.595, binary AUC 0.839, surpassing heavier 3D DenseNet-121 (QWK 0.452, AUC 0.774) on small ordinal cohorts.

6. Select the augmentation ratio ρ via inner cross-validation on the training fold, rather than fixing it globally, to adapt the synthetic-sample injection to the specific combiner and task.

What the improved AI system can do:

• Robustly stratify IPMN risk (no/low/high) with higher accuracy and better ordinal consistency (near-zero two-step misclassifications, 0.02/0.04) than single heavy architectures.

• Automatically choose whether to apply cUPMI based on combiner type (tree vs. linear) and modality fusion (radiomics-only vs. fused), saving compute while maximizing AUC/QWK.

• Generalize better on small medical imaging cohorts (n≈600–700) by regularizing high-capacity meta-learners, reducing overfitting to training folds.

• Provide calibrated per-class probabilities from the stacked ensemble, enabling downstream decision support with confidence thresholds.

• Extend to other multimodal ordinal classification tasks (e.g., disease staging, severity grading) where paired imaging sequences and sub-regional segmentations are available.