Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Upstream Probabilistic Meta-Imputation for Multimodal Pediatric Pancreatitis Classification

Max A. Nelson, Elif Keles, Eminenur Sen Tasci, Merve Yazol, Halil Ertugrul Aktas, Ziliang Hong, Andrea Mia Bejar, Gorkem Durak, Oznur Leman Boyunaga, Ulas Bagci

TL;DR

The paper tackles the challenge of multimodal pediatric pancreatitis classification with limited data by proposing Upstream Probabilistic Meta-Imputation (UPMI), a lightweight augmentation that operates in the meta-feature space prior to a meta-learner. It constructs a 7‑dimensional meta-feature vector from modality-specific predictions and uses fold-local Gaussian mixtures to generate synthetic meta-features, which are combined with real ones to train a Random Forest meta-classifier. On 67 pediatric subjects with paired T1W/T2W MRI, UPMI achieves a mean AUC of $0.908 \pm 0.072$, a ~5% relative improvement over a real-only baseline of $0.864 \pm 0.061$, demonstrating that leakage-aware, decision-layer augmentation can boost performance in small-sample multimodal imaging. The approach is computationally efficient, interpretable, and applicable to other multimodal, small-sample medical prediction tasks.

Abstract

Pediatric pancreatitis is a progressive and debilitating inflammatory condition, including acute pancreatitis and chronic pancreatitis, that presents significant clinical diagnostic challenges. Machine learning-based methods also face diagnostic challenges due to limited sample availability and multimodal imaging complexity. To address these challenges, this paper introduces Upstream Probabilistic Meta-Imputation (UPMI), a light-weight augmentation strategy that operates upstream of a meta-learner in a low-dimensional meta-feature space rather than in image space. Modality-specific logistic regressions (T1W and T2W MRI radiomics) produce probability outputs that are transformed into a 7-dimensional meta-feature vector. Class-conditional Gaussian mixture models (GMMs) are then fit within each cross-validation fold to sample synthetic meta-features that, combined with real meta-features, train a Random Forest (RF) meta-classifier. On 67 pediatric subjects with paired T1W/T2W MRIs, UPMI achieves a mean AUC of 0.908 $\pm$ 0.072, a $\sim$5% relative gain over a real-only baseline (AUC 0.864 $\pm$ 0.061).

Upstream Probabilistic Meta-Imputation for Multimodal Pediatric Pancreatitis Classification

TL;DR

The paper tackles the challenge of multimodal pediatric pancreatitis classification with limited data by proposing Upstream Probabilistic Meta-Imputation (UPMI), a lightweight augmentation that operates in the meta-feature space prior to a meta-learner. It constructs a 7‑dimensional meta-feature vector from modality-specific predictions and uses fold-local Gaussian mixtures to generate synthetic meta-features, which are combined with real ones to train a Random Forest meta-classifier. On 67 pediatric subjects with paired T1W/T2W MRI, UPMI achieves a mean AUC of , a ~5% relative improvement over a real-only baseline of , demonstrating that leakage-aware, decision-layer augmentation can boost performance in small-sample multimodal imaging. The approach is computationally efficient, interpretable, and applicable to other multimodal, small-sample medical prediction tasks.

Abstract

Pediatric pancreatitis is a progressive and debilitating inflammatory condition, including acute pancreatitis and chronic pancreatitis, that presents significant clinical diagnostic challenges. Machine learning-based methods also face diagnostic challenges due to limited sample availability and multimodal imaging complexity. To address these challenges, this paper introduces Upstream Probabilistic Meta-Imputation (UPMI), a light-weight augmentation strategy that operates upstream of a meta-learner in a low-dimensional meta-feature space rather than in image space. Modality-specific logistic regressions (T1W and T2W MRI radiomics) produce probability outputs that are transformed into a 7-dimensional meta-feature vector. Class-conditional Gaussian mixture models (GMMs) are then fit within each cross-validation fold to sample synthetic meta-features that, combined with real meta-features, train a Random Forest (RF) meta-classifier. On 67 pediatric subjects with paired T1W/T2W MRIs, UPMI achieves a mean AUC of 0.908 0.072, a 5% relative gain over a real-only baseline (AUC 0.864 0.061).

Paper Structure

This paper contains 17 sections, 5 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methods
  4. Data Aggregation and Annotation
  5. Radiomics Feature Extraction and Selection
  6. Baseline Logistic Regression Classifiers
  7. Synthetic Meta-Feature Extraction
  8. Synthetic Meta-Feature Generation
  9. Meta-Learner Training with Augmented Data
  10. Results
  11. Ablation Study Results
  12. Statistical Analysis
  13. Synthetic Data Quality Validation
  14. Error Analysis
  15. Conclusion
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Representative cases from the pediatric pancreas MRI dataset showing T1W and T2W sequences. Reference pancreas segmentations are highlighted in red. First two rows (Cases 1-2): healthy control subjects; last two rows (Cases 3-4): pancreatitis patients from our cohort.
  • Figure 2: Multi-modality pipeline for pediatric pancreatitis classification with the proposed method UPMI.
  • Figure 3: Violin plot comparison of real and GMM-sampled meta-feature distributions. Titles report two-sample KS p-values; the borderline significant shift feature is shaded.
  • Figure 4: (a) Mean ROC curve across 5 folds, (b) Aggregated confusion matrix; both correspond to optimal config.
  • Figure 5: Representative misclassifications arranged as a confusion‑matrix grid (rows=reference; columns=prediction) using axial T1W/T2W slices. False negatives typically reflect subtle or atypical pancreatitis, whereas false positives often arise from peripancreatic fat stranding; true cases are shown for context. See aggregate counts in Figure 4b.