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

Attention-based Saliency Maps for Pneumothorax Classification using Vision Transformers

https://doi.org/10.1148/ryai.220187

1.0

A.W. (email: ed.mut@kellow.ordnassela)

Supported by the German Federal Ministry of Health’s Program for Digital Innovations for the Improvement of Patient-Centered Care in Health Care (grant agreement no. 2520DAT920).

Retrospective study using public datasets; institutional review board review and patient informed consent were not required.

Vision Transformer (deit_base_distilled_patch16_224) fine-tuned for multi-label chest radiograph classification; baseline comparison with DenseNet-121. Attention-based saliency via Transformer Multimodal Explainability (TMME) and Grad-CAM for comparison.

Assists pneumothorax detection on chest radiographs by providing classification scores and more interpretable attention-based saliency maps that can reveal model focus and potential biases.

Clinical decision support systems; reader aid for image interpretation and model explainability.

Sensitivity and specificity reported at the threshold maximizing F1 score.

Decision support; provides probabilities and saliency maps to assist radiologists; does not automate final diagnosis.

Exploratory research use on adult chest radiographs to classify pneumothorax (and other thoracic findings: cardiomegaly, consolidation, pleural effusion, atelectasis) and provide attention-based saliency maps; intended for radiology reading environments.

Frontal chest radiographs from public datasets, preprocessed and resized to 224×224 pixels; normalization per ImageNet mean/SD; data augmentation during training.

Model outputs class probabilities and saliency maps; attention-based TMME saliency maps were more useful and consistent than Grad-CAM in reader study and quantitative metrics.

Bias observed toward confounders such as chest tubes influencing pneumothorax predictions; limited training data pooled from public datasets; images resized to 224×224 may reduce detectability of subtle pneumothoraces; limited user study size; segmentation masks largely available only for pneumothorax, limiting precise EHR evaluation for other classes.

Attention-based TMME saliency maps are recommended over Grad-CAM for interpretability in this context; saliency maps can help identify model biases (e.g., chest tube reliance).

Training details reported: ViT fine-tuned with SGD (lr=0.001), batch size 64, 500 epochs with early stopping; images normalized to ImageNet statistics; data augmentations (random affine, translation, scale, horizontal flip); DenseNet baseline fine-tuned with SGD (lr=0.001), batch size 32, 30 epochs with early stopping.