Sep 29

SS 13 - DHI 1: The Digital Revolution in Radiation Oncology: AI Models for Enhanced Patient Care

182 - Predictive Modeling Radiation-Induced Visual Field Deficits for Early Toxicity Detection

08:50am - 09:00am PT

Room 20/21

Presenter(s)

Juliette Thariat, MD, PhD - Centre François Baclesse, Caen, Normandy

T. N. Pham¹, N. Azemar², M. Seraphim¹, J. C. Quintyn³, and J. O. Thariat¹; ¹Radiation Oncology Department, Centre François Baclesse, Caen, Normandy, France, CAEN, France, ²Laboratoire de physique corpusculaire UMR6534 IN2P3/ENSICAEN, Université de Caen- Normandie, Caen, France, ³Department of Ophthalmology, University Hospital of Caen, Caen, France

Purpose/Objective(s): Radiation-induced optic neuropathy (RION) occurs in 4% of extraocular tumors and 14% of intraocular tumors. Current predictive models mostly apply to grade 4 RION, i.e. blindness, at an irreversible stage. We developed a more sensitive predictive model using quantitative visual outcomes (visual acuity and visual field) and dosiomics to enable early prediction of moderate-to-severe RION in patients with head and neck cancers (HNC) or central nervous system (CNS) tumors undergoing pencil beam scanning proton therapy.

Materials/Methods: This prospective cohort of patients with paraoptic HNC/CNS tumors included standardized and systematic ophthalmological assessments (visual acuity, visual field perimetry) at baseline and twice yearly, and dose metrics for the optic nerves, optic chiasm, and retinas. RION was defined as visual field deficit lower than -6 dB (grade 2+; corresponding to moderate damage suggested by Hodapp-Parrish-Anderson criteria using in glaucoma). Patients with visual field deficit above -6 dB at baseline and without recovery during follow-up were included. Model performance was assessed using the area under the curve (AUC) and sensitivity, and results were compared against logistic regression and random forest classifiers. Simulations were conducted to evaluate the impact of predictive variables on RION using odd ratios.

Results: Of 238 patients enrolled, 105 patients (179 eyes) met the inclusion criteria. The median patient age was 58.1 years, with meningiomas (N=46), sinonasal tumors (N=25), and pituitary adenoma (N=15) being the most common diagnoses. Baseline visual deficits ranged from -5.4 dB to 0.8 dB, with a median value of -0.8 dB. The median follow-up period was 5.4 years. Following treatment, 42 of 179 eyes developed grade 2+ RION, with a median onset time of 12.1 months.

Published models demonstrated poor predictive performance for grade 2+ RION (AUC = 0.5, sensitivity = 10%). Random forest models combining quantitative endpoints and dosiomics demonstrated a specificity of 0.8 by AUC, and sensitivity of 45%. The most predictive variables were baseline visual field deficit and the volume of the optic chiasm receiving more than 50 Gy (V50Gy_chiasma). Simulations showed that patients with a baseline visual field deficit between [-3;-6) dB were 5.6 times more likely to develop grade 2 or higher RION than those with a deficit [0;-3) dB. Patients with a V50Gy_chiasma of higher than 50% had a 3.5 times higher risk of developing grade 2 or higher RION compared to those with V50Gy_chiasma of lower than 50%.

Conclusion: RION can be predicted early and at a moderate grade, which, along analysis of ganglion cell layer and retinal nerve fiber thickness optical coherence tomography, and visual evoked potentials in our cohort, opens avenues to better decipher the axonal, myelin or endothelial mechanisms of RION.