CXR-LT 2026 Challenge: Projection-Aware Multi-Label and Zero-Shot Chest X-Ray Classification

TL;DR

This paper introduces a challenge framework for multi-label and zero-shot chest X-ray classification, integrating projection-specific models and novel architectures to improve generalization and robustness.

Contribution

It presents a unified framework with projection-aware models and a dual-branch architecture extending CheXzero for enhanced zero-shot and multi-label classification.

Findings

Effective handling of diverse CXR projections with integrated models.

Novel dual-branch architecture improves zero-shot classification.

Robustness achieved through strong data and test-time augmentations.

Abstract

This challenge tackles multi-label classification for known chest X-ray (CXR) lesions and zero-shot classification for unseen ones. To handle diverse CXR projections, we integrate projection-specific models via a classification network into a unified framework. For zero-shot classification (Task 2), we extend CheXzero with a novel dual-branch architecture that combines contrastive learning, Asymmetric Loss (ASL), and LLM-generated descriptive prompts. This effectively mitigates severe long-tail imbalances and maximizes zero-shot generalization. Additionally, strong data and test-time augmentations (TTA) ensure robustness across both tasks.