Fig. 1.

Experiments performed in the current study. Experiment A: Radiological/AD label detection based on demographic attributes via logistic regression. Experiment B: Demographic attribute prediction from CXR and brain MRI images via the CNN-based model. Detection performance was used to indicate the presence of demographic features in CXR or brain MRI images, which could potentially be used as shortcuts. Experiment C: Testing for disparities in radiological/AD label detection results among demographic groups when applying a Densenet121-based model to CXR images and a ResNet 18-based model to brain MRI images. Experiment D (Task transfer test): The trained model in Experiment C would be frozen and the last prediction layer would be replaced to classify demographic attributes. The model's performance was used to indicate whether the model had incorporated demographic features as shortcuts in the radiological/AD label detection task. The proposed augmentation method was then applied to Experiment B–D and compared with the results obtained without augmentation.