Symmetric Siamese Networks for Longitudinal Chest Radiograph Change Detection: A Leakage-Controlled Study on MIMIC-CXR.

Background/

Objectives

Radiologists usually read a chest radiograph by comparing it with an earlier one, yet most deep learning models are trained and tested on single images. This study measures how much a longitudinal pair of radiographs improves finding-specific change detection on MIMIC-CXR and tests whether the gain can be explained by encoder pretraining.

Methods

One frontal pair per patient was built from the earliest and latest studies within 180 days (14,043 pairs), and paired CheXpert labels defined four transition classes: absent-to-absent, onset, resolved, and persistent. A Siamese DenseNet-121 with two symmetric, weight-sharing branches fused the two images through concatenation and feature difference, and it was compared with a matched baseline that received only the final image. Evaluation used patient-grouped five-fold cross-validation over five seeds and three leakage-free encoder regimes.

Results

On two matched binary tasks that hold the final image fixed-resolved versus absent (1→0 vs. 0→0) and onset versus persistent (0→1 vs. 1→1)-the single-image baseline performed at chance (AUROC 0.45 to 0.60), while the paired model scored well above it (paired difference +0.18 to +0.37 for resolution and +0.12 to +0.34 for onset; all significant by patient-level bootstrap after multiple-comparison correction), reaching a resolved-versus-absent AUROC of up to 0.87 (support devices). The benefit was stable across seeds, encoder regimes, projection changes, and follow-up intervals, and the CheXpert-derived transitions agreed with radiologist annotations (Cohen's κ up to 0.97).

Conclusions

A prior radiograph adds value whenever the final image alone underdetermines the transition, and they support pair-based designs for longitudinal change analysis.