Curriculum-guided AI framework boosts semi-supervised chest X-ray disease classification

A machine learning framework called Curriculum-guided Adaptive Representation Decoupling, or CARD, achieved a Macro-AUROC of 75.9% for semi-supervised chest disease classification from chest X-rays, outperforming the STiL comparison method by 1.1 percentage points under a matched radiographic protocol. When matched pre-diagnostic clinical text was added to the radiographic inputs, the Macro-AUROC rose to 77.0%. The study, published in Medical & Biological Engineering & Computing on 24 October 2026, was conducted by Suli Wang and Shengyong Ding.

CARD was designed for a setting in which automated chest X-ray diagnosis is constrained by annotation scarcity and heterogeneous clinical information. Frontal and lateral radiographs provide complementary views within one X-ray modality, while pre-diagnostic clinical context constitutes a heterogeneous textual modality. The framework combines these sources in a multimodal semi-supervised pipeline intended to classify chest diseases despite limited labeled data.

The first component, Label-aware Dynamic Curriculum Learning, partitions unlabeled sample-label pairs into easy, stable, and hard groups. Rather than treating all unlabeled samples equally, the method combines prediction persistence with dynamic confidence thresholds to decide how reliable each self-generated label is. This curriculum-style organization allows the model to draw on the most dependable pseudo-labeled examples first and to incorporate more ambiguous ones progressively, instead of weighting every self-generated label identically.

A second component, Persistence-guided Adaptive Feature Decoupling, transfers these reliability cues to cross-view representation learning. Chest X-ray examinations include frontal and lateral projections, which capture overlapping but not identical radiographic information. The framework balances view-consistent alignment against the preservation of projection-specific radiographic information, so that features shared by both views can be aligned while details unique to each projection are retained. The learned multi-view radiographic representation is then fused with clinical-text features for the final multimodal prediction.

The authors present the reported results as support for reliability-aware curriculum learning, adaptive cross-view representation decoupling, and clinically aligned multimodal fusion in chest disease classification. The measured performance figures come from the study’s matched radiographic protocol and its multimodal extension, and the reported gains are expressed in Macro-AUROC percentage points rather than as clinical outcomes in patients.

Subject of Research: A curriculum-guided multimodal semi-supervised deep learning method for chest disease classification from chest X-rays and clinical text

Article Title: Curriculum-guided adaptive representation decoupling for multimodal semi-supervised chest disease classification

Article References: Wang, S., & Ding, S. (2026). Curriculum-guided adaptive representation decoupling for multimodal semi-supervised chest disease classification. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03697-8

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03697-8

Keywords: chest X-ray, semi-supervised learning, multimodal fusion, curriculum learning, medical imaging, disease classification, Macro-AUROC, clinical text, machine learning, Curriculum-guided, adaptive, representation