A Machine Learning-Based Approach for Classifying Optic Nerve Images Associated with Papilledema

Idiopathic intracranial hypertension (IIH) can present with papilledema, and delayed recognition of optic disc swelling may threaten vision. This exploratory study evaluates whether a classical machine-learning pipeline can distinguish fundus images carrying repository-provided labels of normal, papilledema, and pseudopapilledema.

The public OSF archive contained 1,369 images: 779 normal, 295 papilledema, and 295 pseudopapilledema. Images were converted to 8-bit grayscale, resized to 64 x 64 pixels, scaled to [0, 1], and flattened. A stratified image-level 80:20 split (random seed 42) produced 1,095 training and 274 test images. L2-regularized logistic regression was compared with a class-prior dummy classifier, 5-nearest neighbors, and a 300-tree random forest.

Logistic regression achieved 87.2% accuracy, macro precision 0.842, macro recall 0.844, macro F1 0.840, and macro one-vs-rest AUC 0.959. It outperformed the dummy classifier (56.9% accuracy), 5-nearest neighbors (83.9%), and random forest (84.7%). Pseudopapilledema had the lowest recall (0.712), with 11 of 59 images classified as papilledema.

The results provide a reproducible image-level benchmark, not evidence of patient-level or external clinical generalization. The archive contains no patient identifiers, so a patient-grouped split could not be performed. In addition, the provenance of the repository's papilledema folder requires caution because the source publication describes a broader optic-neuropathy group. Patient-linked, clinically adjudicated, multi-center data are needed before clinical use.