Establish genuine clinical agreement between CMTN and radiologist judgment

Establish whether the Cross-Modal Triage Network (CMTN) can achieve clinically valid severity-triage agreement with independent radiologist judgment rather than merely reproducing severity labels generated by the CheXpert natural-language-processing pipeline.

Background

The CMTN achieved high agreement with automatically derived reference labels, but its agreement with a blinded independent radiologist in the Phase I audit was substantially lower. The paper explains that the automated labels are extracted from finalized radiology reports and therefore may not represent genuine clinical severity judgments.

Because the audit included only 100 cases and a single radiologist, the authors state that the model has not yet demonstrated genuine clinical agreement. Resolving this issue requires larger, multi-reader clinical validation using radiologist-derived labels before prospective deployment.

References

The Phase I audit is the only evaluation in this study that measures agreement with authentic clinical judgment directly, and it indicates that CMTN has not yet demonstrated this capability.

— Cross-modal triage network: a multimodal deep learning framework for severity-based triage and visual explainability in chest radiographs  (2609.04357 - Ghulam et al., 3 Sep 2026) in Section 4, Discussion