Clinically-Informed Modeling for Pediatric Brain Tumor Classification from Whole-Slide Histopathology Images

1. Problem Statement

The paper addresses the critical challenge of diagnosing pediatric brain tumors using Whole-Slide Images (WSIs) via deep learning. Unlike adult oncology, pediatric brain tumor diagnosis faces unique hurdles:

• Data Scarcity: Pediatric cases are rare, leading to small cohort sizes (often <300 patients) even at large institutions, unlike the tens of thousands available for adult cancers.
• Class Imbalance & Fine-Grained Overlap: The WHO classification defines many rare subtypes with significant histopathological overlap on H&E slides. Distinguishing between diagnostically confusable subtypes (e.g., low-grade gliomas vs. pilocytic astrocytomas) is difficult.
• Weak Supervision: Detailed patch-level annotations are rarely available. Models must rely on slide-level labels via Multiple Instance Learning (MIL), but standard MIL often fails to learn robust geometries for fine-grained separation in low-data regimes.
• Limitations of Foundation Models: While pathology foundation models (like UNI) provide strong patch-level features, their downstream adaptation for slide-level classification in data-scarce, fine-grained settings remains underexplored.

2. Methodology

The authors propose an Expert-Guided Contrastive Learning (EGCL) framework integrated into a slide-level MIL pipeline.

A. Pipeline Architecture

1. Feature Extraction: WSIs are tiled into 256×256 patches. A frozen UNI2-h (ViT-H/14) foundation model, pre-trained on 200M+ histology patches, extracts patch embeddings.
2. Slide Aggregation: The CLAM (Clustering-constrained Attention MIL) architecture aggregates patch embeddings into a single slide representation. It uses class-specific attention mechanisms to weight informative patches and imposes clustering constraints.
3. Loss Function: The total loss combines:
   • $L_{bag}$: Cross-entropy for slide-level classification.
   • $L_{inst}$: Instance-level SVM loss for attention supervision.
   • $L_{CL}$: Contrastive Loss (InfoNCE style) applied to the slide-level embeddings.
   • $L_{reg}$: L2 regularization.

B. Contrastive Learning Strategies

The core innovation lies in how the contrastive loss is applied during fine-tuning:

• Standard Supervised Contrastive Learning (CL): Encourages embeddings of slides with the same label to cluster together and different labels to separate, using a memory queue of negative samples.
• Expert-Guided Contrastive Learning (EGCL): A novel variant where the negative sampling is biased by clinical knowledge.
  • Hard Negatives: Experts define pairs of classes that are clinically confusable (e.g., Normal Brain vs. Non-neoplastic lesions, or Pilocytic Astrocytoma vs. Ependymal tumors).
  • Mechanism: During training, the contrastive denominator is up-weighted for these specific "hard negative" pairs. This explicitly forces the model to learn finer decision boundaries between morphologically similar subtypes, rather than just separating broad categories.

C. Experimental Setup

• Dataset: 237 de-identified pediatric brain WSIs from Dell Children's Medical Center (7 diagnostic categories).
• Tasks: Evaluated across four granularities:
  • Task 1: Binary (Tumor vs. Non-tumor).
  • Task 2: 7-Class (Full spectrum).
  • Task 3: 6-Class (Merging Normal/Non-neoplastic).
  • Task 4: 3-Class (Fine-grained: DMG H3-mutated, Pilocytic Astrocytoma, Ependymal).
• Validation: Patient-stratified 10-fold cross-validation to prevent data leakage.

3. Key Contributions

1. Expert-Guided Contrastive Framework: Introduced EGCL, which integrates clinical domain knowledge directly into the loss function by prioritizing "hard negatives" (confusable subtypes) during contrastive fine-tuning.
2. Slide-Level Geometry Regularization: Demonstrated that explicitly shaping the geometry of slide-level representations via contrastive loss is crucial for fine-grained classification in data-scarce settings, going beyond standard MIL aggregation improvements.
3. Comprehensive Benchmarking: Provided a systematic comparison of foundation-model-based MIL baselines, general contrastive learning, and expert-guided variants across multiple classification granularities.
4. Clinical Error Analysis: Conducted a detailed error analysis with neuropathologists to interpret model failures and successes, revealing that contrastive learning helps overcome "superficial mimics" but can sometimes over-weight specific features like cellularity.

4. Results

• Performance Gains:
  • Fine-Grained Tasks (7-Class & 3-Class): Contrastive learning yielded measurable improvements. In the 7-class task, CL + CLAM achieved the best Macro Recall (0.768), significantly outperforming the Baseline (0.705).
  • Specific Subtypes: EGCL showed particular strength in separating difficult pairs, achieving high recall for Ependymal tumors (0.900) and High-grade tumors (0.850) in the 6-class setting.
  • Binary Task: No significant gain was observed in the binary (Tumor vs. Non-tumor) task, as the decision boundary is coarse and already near-ceiling performance.
• Hyperparameter Sensitivity: The contrastive loss weight ($\lambda$) followed an inverted-U trend. An intermediate weight (e.g., $\lambda=0.5$) provided optimal regularization, while excessive weights ($\lambda > 0.8$) distorted decision boundaries.
• Stain Normalization: Macenko stain normalization showed mixed results; it helped in binary tasks but did not consistently improve fine-grained classification, suggesting the UNI2-h foundation model is robust to color variations.
• Error Analysis:
  • Successes: CL/EGCL corrected "superficial mimics" where the baseline confused non-glial tumors with low-grade gliomas due to similar cellular density.
  • Failures: In some cases, CL/EGCL over-weighted "cellularity," misclassifying a dense low-grade tumor as high-grade. Some cases remained ambiguous across all models due to intrinsic histological overlap (Bayes error), highlighting the need for molecular data.

5. Significance

This work highlights that data scarcity in pediatric pathology cannot be solved by larger models alone; the geometry of the learned representations must be explicitly shaped by domain knowledge.

• Clinical Impact: By improving the separation of diagnostically confusable subtypes, the framework aids in more accurate risk stratification and treatment planning for children.
• Methodological Insight: It establishes that "hard negative" mining guided by clinical experts is a powerful strategy for weakly supervised learning in medical imaging, particularly when datasets are too small for standard deep learning to learn these subtle distinctions from scratch.
• Future Direction: The study underscores the necessity of integrating molecular/genetic data or multi-modal inputs to resolve cases where histology alone is inherently ambiguous.