What does this research mean for the field?

Deep learning models utilizing longitudinal pairs of chest radiographs significantly outperform single-image models in detecting finding-specific clinical changes. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To determine how much incorporating prior longitudinal chest radiographs improves finding-specific change detection over single-image deep learning models and verify if gains stem from architectural design rather than encoder pretraining.

September 10, 2026TomographyOpen Access

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

View Full Paper

Ask AI

Bookmark

Authors

ŞIŞahin Işık HEHakan Alp Eren

Discussion

Member takes

Overview

Retrospective validation study demonstrates superior change detection using longitudinal paired chest radiographs over single-image models, highlighting the critical value of prior imaging data.

Key Points

To determine how much incorporating prior longitudinal chest radiographs improves finding-specific change detection over single-image deep learning models and verify if gains stem from architectural design rather than encoder pretraining.
Extracted 14,043 frontal radiograph pairs (earliest and latest within 180 days per patient) from MIMIC-CXR, labeled into four transitions: absent-to-absent, onset, resolved, and persistent.
Trained a symmetric, weight-sharing Siamese DenseNet-121 fusing image pairs via concatenation and feature difference, evaluated against a single-image baseline using patient-grouped five-fold cross-validation across five seeds.
Single-image models performed near chance on matched transition tasks holding the final image fixed (AUROC 0.45 to 0.60).
The paired Siamese model outperformed the baseline with significant AUROC gains of +0.18 to +0.37 for resolution and +0.12 to +0.34 for onset (patient-level bootstrap), reaching an AUROC up to 0.87 for support devices.

Cite This Study

Işık et al. (2026) studied this question.

synapsesocial.com/papers/6aa27c3058559d80afc75e37 https://doi.org/10.3390/tomography12090129

View Full Paper

Ask AI

Bookmark

September 10, 2026TomographyOpen Access

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

View Full Paper

Ask AI

Bookmark

Authors

ŞIŞahin Işık HEHakan Alp Eren

Discussion

Member takes

Overview

Retrospective validation study demonstrates superior change detection using longitudinal paired chest radiographs over single-image models, highlighting the critical value of prior imaging data.

Key Points

To determine how much incorporating prior longitudinal chest radiographs improves finding-specific change detection over single-image deep learning models and verify if gains stem from architectural design rather than encoder pretraining.
Extracted 14,043 frontal radiograph pairs (earliest and latest within 180 days per patient) from MIMIC-CXR, labeled into four transitions: absent-to-absent, onset, resolved, and persistent.
Trained a symmetric, weight-sharing Siamese DenseNet-121 fusing image pairs via concatenation and feature difference, evaluated against a single-image baseline using patient-grouped five-fold cross-validation across five seeds.
Single-image models performed near chance on matched transition tasks holding the final image fixed (AUROC 0.45 to 0.60).
The paired Siamese model outperformed the baseline with significant AUROC gains of +0.18 to +0.37 for resolution and +0.12 to +0.34 for onset (patient-level bootstrap), reaching an AUROC up to 0.87 for support devices.

Cite This Study

Işık et al. (2026) studied this question.

synapsesocial.com/papers/6aa27c3058559d80afc75e37 https://doi.org/10.3390/tomography12090129

View Full Paper

Ask AI

Bookmark