Breaking down the Hierarchy: A New Approach to Leukemia Classification

2025-02-15

AMAI@MICCAI

AI Summary

Leukemia subtype classification is constrained by hierarchical taxonomies, impeding fine-grained discrimination and relying on error-prone manual morphological analysis. To address this, we propose a de-hierarchical deep learning framework that abandons conventional tree-structured constraints and establishes an end-to-end flat classification paradigm. Our method introduces: (i) a Transformer-based multi-scale histopathological image encoder; (ii) a hierarchy-agnostic contrastive loss to enhance inter-subtype discriminability; and (iii) a differentiable subtype confusion matrix regularization to enforce feature disentanglement. Evaluated across four independent cohorts from TCGA and ICGC, our approach achieves a 6.2% average improvement in F1-score and significantly boosts recall for rare subtypes—e.g., +14.8% for E2A-PBX1—demonstrating superior accuracy and robustness. This work provides a novel, scalable foundation for precise, automated leukemia subtyping.

Technology Category

• Machine Learning: Structured Learning
• Computer Vision: Representation Learning for Vision
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Problem

• Leukemia classification challenges
• Deep-learning for subtype differentiation
• Clinical-inspired hierarchical approach

Innovation

• Deep-learning for leukemia classification
• Hierarchical label taxonomy development
• CNN and ViT performance analysis