Open Access

ARTICLE

Disease-Aware Multi-Relational Representation Learning and Neighborhood-Prior Decision Calibration for Chest X-Ray Multi-Label Prediction

Yuxin Zhang^1,#, Bin Li^1,#,*, Riqiang Liao², Lianfang Tian¹

1 School of Automation Science and Engineering, South China University of Technology, Guangzhou, China
2 Guangdong Lung Cancer Institute, Guangdong Provincial People’s Hospital & Guangdong Academy of Medical Sciences, Guangzhou, China

* Corresponding Author: Bin Li. Email:
# These authors contributed equally to this work

(This article belongs to the Special Issue: Emerging Artificial Intelligence Technologies and Applications-II)

Computer Modeling in Engineering & Sciences 2026, 148(3), 43 https://doi.org/10.32604/cmes.2026.087646

Received 24 June 2026; Accepted 01 September 2026; Issue published 28 September 2026

Abstract

Chest X-ray multi-label prediction aims to identify multiple thoracic findings from radiographs and generate structured disease information for clinical analysis. Conventional image-only supervised learning uses disease labels primarily as prediction targets and may underutilize the cross-sample relational structure contained in multi-label disease annotations, thereby limiting disease-aware representation learning. Structured disease-label-guided contrastive pre-training provides a promising solution by encoding report-derived disease-label vectors as an auxiliary training view and aligning them with image representations through contrastive learning, while preserving image-only inference. In this formulation, the binary disease-label vector serves as both the structured input to the auxiliary encoder and the downstream prediction target, allowing its relational structure to guide image representation learning without introducing additional patient-level clinical variables. However, two major challenges remain. First, conventional instance-level contrastive alignment treats only the matched image and disease-label representation as a positive pair and regards other samples as negatives, which may introduce false-negative supervision when different patients share related disease combinations or label co-occurrence patterns. Second, fixed-threshold decision rules ignore label imbalance and patient-specific disease tendencies when converting continuous probabilities into binary labels. To address these limitations, we propose a two-stage framework for chest X-ray multi-label prediction. During pre-training, Multi-Relational Cross-Modal Alignment (MRCA) constructs clinical multi-relation targets from the paired identity relation, disease-label similarity, and tabular phenotype similarity, and converts them into soft target distributions for bidirectional soft alignment between image and encoded disease-label representations. This design reduces inappropriate repulsion between non-paired samples with related disease-label patterns and encourages the image encoder to preserve disease-aware inter-sample relations. During prediction, Neighborhood-Prior Decision Calibration (NPDC) converts continuous disease probabilities into binary labels by estimating sample-label-specific thresholds and refining them with local neighborhood disease priors retrieved from an embedding-based training memory bank. Experiments using CheXpert development data with held-out CheXlocalize testing, together with independent experiments on NIH Chest X-rays, demonstrate that MRCA improves disease-aware representation learning and score-level discrimination, while NPDC provides a balanced threshold-dependent prediction profile by jointly considering sample-level consistency and label-wise decision quality. Under the CheXpert evaluation protocol, the proposed framework achieves the best Exact Match, Macro-AUROC, Macro-AUPRC, and Macro-F1 among the main comparison strategies. On NIH Chest X-rays, it obtains the highest Macro-AUROC, Macro-AUPRC, and Macro-F1, demonstrating consistent effectiveness on an additional dataset with imbalanced multi-label distributions.

---

Keywords

---

Supplementary Material

Supplementary Material File

---