Siamese Networks Improve Longitudinal Chest Radiograph Change Detection on MIMIC-CXR

On MIMIC-CXR, a Siamese network using paired chest radiographs outperformed a single-image model for detecting resolved vs absent findings (AUROC up to 0.87 vs chance 0.45-0.60), demonstrating the value of prior images for longitudinal change detection.

• CheXpert-derived transition labels (absent-to-absent, onset, resolved, persistent) agreed well with radiologist annotations, achieving Cohen’s κ up to 0.97.
• The Siamese model’s benefit remained stable across different encoder pretraining regimes, projection changes, and follow-up intervals in a patient-grouped cross-validation design.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.