Artifact-based Domain Generalization of Skin Lesion Models

This is the official repository of the paper "Artifact-based Domain Generalization of Skin Lesion Models", accepted at the ISIC Workshop @ ECCV 2022.

Reproducing our results:

Data

The training/validation/test data is passed through two specific parameters:

{train|val|test}_csv : is a csv containing the list of samples on the set. On folder trap_sets, we include all the csvs used in the work, which are based on ISIC 2019.

root_dir: is the directory where samples can be found. Alternatively, it is possible to include the full path on the csvs mentioned above.

The confounder annotation is at the file, which is referenced in the code at

artifact-generalization-skin/data/skin_dataset.py

Line 27 in ce89fef

self.attrs_df = pd.read_csv('/artifact-generalization-skin/isic_inferred_wocarcinoma.csv')

For running the out-of-distribution evaluation, include images on the folder datasets. They are loaded at

artifact-generalization-skin/train.py

Lines 277 to 280 in ce89fef

	test_ds_atlas_clin = CSVDataset('/artifact-generalization-skin/datasets/edraAtlas', '/artifact-generalization-skin/datasets/edraAtlas/atlas-clinical-all.csv', 'image', 'label', transform=get_transform_skin(False), add_extension='.jpg')
	test_ds_atlas_derm = CSVDataset('/artifact-generalization-skin/datasets/edraAtlas', '/artifact-generalization-skin/datasets/edraAtlas/atlas-dermato-all.csv', 'image', 'label',transform=get_transform_skin(False), add_extension='.jpg')
	test_ds_ph2 = CSVDataset('/artifact-generalization-skin/datasets/ph2images/', '/artifact-generalization-skin/datasets/ph2images/ph2.csv', 'image', 'label',transform=get_transform_skin(False), add_extension='.jpg')
	test_ds_padufes = CSVDataset('/artifact-generalization-skin/datasets/pad-ufes/', '/artifact-generalization-skin/datasets/pad-ufes/padufes-test-wocarc.csv', 'img_id', 'label',transform=get_transform_skin(False), add_extension=None)

General

• The code is fully prepared to use wandb, but it is disabled by default.
• We make use of the sacred library, allowing the organization of the results by folder, according to the name passed in the parameter exp_name.
• We make available the script to run all trainings and evaluations.