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

2026-04-24 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

A new expert-guided contrastive fine-tuning framework has been developed for pediatric brain tumor diagnosis from whole-slide histopathology images (WSIs). This approach addresses challenges like data scarcity, class imbalance, and fine-grained morphological overlap by integrating contrastive learning into slide-level multiple instance learning (MIL). The framework uses a frozen UNI2-h pathology foundation model for patch-level feature extraction and a multi-branch CLAM architecture for slide-level classification. It includes both a general supervised contrastive setting and an expert-guided variant that incorporates clinically informed hard negatives targeting diagnostically confusable subtypes. Experiments on a 237-WSI dataset from Dell Children's Medical Center, across 2-, 3-, 6-, and 7-class tasks, demonstrate that contrastive fine-tuning, particularly in multi-class settings, measurably improves fine-grained diagnostic distinctions and representation geometry, with optimal performance at intermediate contrastive loss weights.

Key takeaway

For Computer Vision Engineers developing diagnostic tools for pediatric pathology, integrating expert-guided contrastive learning into your WSI classification pipelines can significantly improve fine-grained tumor subtype differentiation, especially in data-scarce scenarios. You should experiment with intermediate contrastive loss weights to balance regularization and classification accuracy, and consider incorporating clinical knowledge to define hard negative pairs for better class separation.

Key insights

Contrastive learning, especially with expert-guided hard negatives, improves fine-grained pediatric brain tumor classification from WSIs.

Principles

• Explicitly shaping slide-level representations enhances fine-grained classification.
• Clinically informed hard negatives improve inter-class separation.
• Optimal contrastive loss weight prevents excessive regularization.

Method

The method involves segmenting WSIs into patches, extracting 1536-dimensional features using a frozen UNI2-h model, and feeding these into a multi-branch CLAM architecture. A queue-based supervised contrastive loss, with or without expert-guided hard negatives, regularizes slide-level representations during MIL fine-tuning.

In practice

• Use UNI2-h as a robust, stain-invariant patch feature extractor.
• Implement patient-stratified cross-validation for robust evaluation.
• Consider contrastive loss for fine-grained multi-class medical image tasks.

Topics

• Pediatric Brain Tumor Classification
• Whole-Slide Images
• Contrastive Learning
• Multiple Instance Learning
• Pathology Foundation Models

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Multi-Class Brain Tumor Classification Using Advanced Deep Learning Models: A Comparative Study — EfficientNetB0 significantly improves multi-class brain tumor classification from MRI images, especially for subtle meningiomas, outperforming other CNNs.
• DaX: Learning General Pathology Representations Across Scales — DaX is a pathology vision foundation model using DINOv3-style self-supervised learning for robust, transferable representations across diverse histopathology tasks.
• Contrast-Informed Augmentation and Domain-Adversarial Training for Adult-to-Neonatal MR Reconstruction Generalization — Contrast-informed augmentation and domain-adversarial training enhance deep learning MR reconstruction generalization from adult to neonatal data.
• CA-GCL: Cross-Anatomy Global-Local Contrastive Learning for Robust 3D Medical Image Understanding — CA-GCL uses global-local contrastive learning and clinical-aware text augmentation to prevent representation collapse in 3D medical FVLP.
• Cascade Classification of Dermoscopic Images of Skin Neoplasms with Controllable Sensitivity and External Clinical Validation — Cascade classification with a tunable triage threshold offers superior sensitivity control and better mimics clinical diagnosis logic.
• Improving clinical interpretability of linear neuroimaging models through feature whitening — A novel whitening method improves neuroimaging model interpretability by decorrelating anatomically informed brain regions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.