Connecting Visual Perception With Proton Therapy-Induced Optic Damage Using Archetypal Analysis

Radiotherapy can lead to radiation-induced optic neuropathy, with vision loss and visual field deficits related to localized damage to the optic pathways. Accurately quantifying visual field deficits and establishing their spatial relationship with anatomical structures and radiation dose distribution remains a significant challenge. We applied archetypal analysis of visual fields as a novel artificial intelligence approach in oncology to identify distinct, interpretable patterns of visual field loss and to model their spatial evolution over time after pencil beam scanning proton therapy.

Machine learning of standardized automated static visual field perimetry was used to decompose the high-dimensional visual field data into convex combinations of representative visual loss patterns, that is, visual archetypes, at both the eye and patient levels. Associations between archetype proportions and radiation dose metrics were evaluated using linear regression, stratified by baseline visual field to account for pre-existing deficits.

In 236 patients, 7 archetypal patterns of visual field loss were identified, including tunnel vision, temporal hemianopia, and diffuse full-field loss. More severe patterns were more frequently observed in patients with meningiomas and pituitary adenomas. Longitudinal analysis revealed an annual reduction of 1.9% in the normal visual field archetype in patients with no-to-mild baseline deficits and 9.7% in those with moderate deficits. Dose-archetype associations were clinically significant after adjusting for baseline deficits. Each 1 Gy increase in minimum chiasm dose was associated with a 0.2% decrease in the normal vision archetype.

This is the first application of archetypal analysis for predicting radiation-induced optic neuropathy. It enables spatially grounded reconnection between patient perception and radiation damage along optic pathways. This approach offers new mechanistic insights into optic pathway injury towards voxel-level correlation between radiation dose and functional loss and supports data-driven personalization of radiotherapy.