Extended Data Fig. 3: Experiment results showing short-cut learning successfully detected by CAML.

Cases revealing classifier shortcut learning behavior through CAML. a. Class-associated manifold from the PALM dataset. b. A series of images generated along the path, with classifier (trained on the PALM dataset where pathological myopia samples undergo brightness enhancement) predictions displayed above each image (values in parentheses indicate the predicted probability of pathological myopia). During counterfactual generation process, only brightness changes in the image, yet the classifier’s prediction shifts, indicating that the classifier uses brightness as a shortcut feature for pathological class identification.