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

BERT models for device labeling in chest radiograph reports

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9344209/

1.0

Jesse C. Rayan; email: ude.nretsewhtuostu@nayar.essej

Authors declared no funding for this work.

IRB approved; designated exempt status with waived informed consent; HIPAA-compliant.

Transformer-based language models (BERT, PubMedBERT, DistilBERT, RoBERTa, DeBERTa) fine-tuned for multi-label binary text classification (presence/absence of devices) using PyTorch Lightning/PyTorch.

Not stated.

Rapid, accurate autonomous annotation of large radiology report datasets for presence/absence of support devices to facilitate downstream computer vision model development and workflow efficiency.

Workflow optimization; data curation/annotation for AI development.

Fully automated inference for report labeling; training requires labeled examples.

Annotating adult chest radiograph text reports to indicate presence or absence of endotracheal tubes, enterogastric tubes, central venous catheters, and Swan-Ganz catheters in an academic radiology setting.

Free-text adult chest radiograph reports (April 2020–March 2021) from a single academic center; 1004 reports manually labeled for training/validation/testing; full corpus 69,095 reports for inference timing.

Fine-tune pretrained transformer model on institution-specific labeled reports (12 epochs; AdamW; initial LR 5e-6 with warmup to 5e-5 by epoch 2 and cosine annealing to 1e-6 by epoch 12). Use fivefold cross-validation with stratification; select best epoch by lowest validation loss; apply fixed weights for test and large-scale inference.

Single-center dataset; mix of structured/unstructured reports; potential negative set bias (64.4% without devices); relatively few SGC-positive cases; known failure modes include confusion of context (e.g., not recognizing “removal” phrasing) and misinterpretation of tube-related terms; unknown generalizability to other institutions and demographics; terminology limited to those observed in 1004 annotated reports.

Use pretrained, domain-specific or newer transformer models (e.g., PubMedBERT, RoBERTa, DeBERTa) to achieve high performance with small training sets for report annotation.

Fivefold cross-validation with stratified folds; best-epoch selection by validation loss; detailed training regimen and hardware specified. External validation not performed.

Training/validation times ranged from 3m39s (DistilBERT) to 22m48s (DeBERTa) on a single Nvidia GTX 1080 Ti GPU; large-scale inference on 69,095 reports completed in as little as 6m15s.