https://atlas.rsna.org/schemas/2025-11/model.json

Twin AGXNet for Anatomy-specific Progression Classification in Chest Radiographs

https://dx.doi.org/10.1148/ryai.230277

1.0

ude.ub@namtab

Pennsylvania Department of Health (grant 4100087331); National Institutes of Health (R01HL141813); National Science Foundation (1839332); computational resources: Bridges-2 system at Pittsburgh Supercomputing Center (NSF award OAC-1928147).

Retrospective study using publicly available MIMIC-CXR dataset; institutional review board approval and patient informed consent were not required.

Twin (siamese-style) Anatomy-Guided chest x-ray Network (AGXNet) with two branches (anatomy and observation networks) linked by a residual attention module; each branch uses DenseNet121 backbone with ImageNet initialization; pretraining with 69 weak labels, fine-tuning on paired images for four-class progression classification.

Automated monitoring of interval changes and detection of new pathologic conditions on chest radiographs to support timely clinical interventions.

Clinical decision support systems; potential alerting of new or progressing findings during image interpretation.

Assistive decision support; provides classification and localization outputs to aid but not replace clinician judgment.

Classification of progression (improved, unchanged, worsened, new) for six observations—atelectasis, consolidation, edema, effusion, pneumonia, pneumothorax—on pairs of frontal chest radiographs from the same patient within one year; intended for hospital settings with chest radiography and associated reports.

Two consecutive frontal chest radiographs from the same patient and the referenced anatomic landmark (when available) derived from prior report; images resized to 512×512.

Model is pretrained on single images with weak labels and fine-tuned on consecutive image pairs; during inference for improved/unchanged/worsened classes, use anatomy CAMs for conditioning; for 'new' class, use constant attention in place of unknown CAMs.

Trained and evaluated only on MIMIC-CXR (primarily ICU patients); terminology for progression extraction not exhaustive; only frontal radiographs used; variability from acquisition factors (positioning, inspiration) not adjusted; progression modeled categorically and pairwise (not quantitative over multiple studies); no external dataset evaluation; performance metrics based on five random splits and paired t tests; manual verification limited to sampled labels.

Further validation on external datasets and non-ICU populations is necessary before clinical deployment.

Five random 70:10:20 data partitions with consistent pretraining/fine-tuning splits and no patient overlap; model selection by validation loss/accuracy; optimization details and software versions reported (AdamW, SciPy 1.13.1).