basedrhys / non-roi-masking

Evaluation of non-ROI masking to improve OOD generalization in chest x-ray disease classification

An Emperical Evaluation of Reducing Spurious Signals in Chest X-Rays via Non-ROI Masking

Weights and Biases Project

Full Report

Contents

• 0-seg_train.ipynb - Train the lung segmentation model
• 1-seg_apply.py - Create masks for a classification dataset via the segmentation model
• 2-smooth_masks.py - Postprocess the predicted lung masks to smooth them out
• 3-create_chexpert.ipynb - Create the binary classification version of CheXpert
• 3-create_datasets.py - Apply the smoothed masks and create train/val/test splits of classification dataset
• 4-clf_train.ipynb - Train the downstream classification model
• 4-clf_train.py - Train the downstream classification model
• 5-eval.py - Evaluate the trained classification models