Automated grading of fundus images could expand screening for diabetic retinopathy, but models trained on one source often degrade sharply when applied to new cameras, clinics, or patient populations. The new arXiv preprint, CF-NODE, proposes a counterfactual spatial node intervention for test-time adaptation, aiming to adjust the model's internal representations during inference rather than retraining.
The method was evaluated across five diabetic retinopathy datasets, according to the abstract. The abstract also references existing test-time adaptation work, but the excerpt ends before providing details. The paper is listed as new on arXiv under identifier 2610.22437.