Machine learning prediction-informed gaze optimization in ocular proton therapy with NTCP evaluation

Ocular proton therapy (OPT) treatment planning has remained largely static since the introduction of the EyePlan system over four decades ago, which is still widely used across treatment centers. The current method relies on planners individually selecting gaze fixation points to optimize eye orientation, aiming to minimize the exposure to sensitive sensitive ocular organs at risk (OARs). However, this manual process lacks explicit information on how eye orientation correlates with toxicity risks. This limitation presents an opportunity to enhance planning efficiency, reduce planner biases, and provide toxicity-informed guidance through automated computational tools.

The primary objective of this research is to develop a computationally efficient automated approach for OPT treatment planning that optimizes eye orientation during treatment. By providing intuitive and clinically relevant information to treatment planners, the approach aims to minimize healthy tissue exposure and reduce the risk of radiation-induced toxicities.

Dose calculations are conducted using the in-house developed treatment planning system (OCULARIS) guided by machine learning (ML) gaze predictions and a tailored cost-function that allows to quickly identify eye orientations with favorable healthy tissue sparing suitable for treatment. The orientations identified by the optimization algorithm are subsequently evaluated against the clinically defined orientations, utilizing both cost metrics and recent normal tissue complication probability (NTCP) models, acknowledging the inherent uncertainties associated with these models. The resulting plans are then validated and evaluated by two clinically experienced medical physicists.

In 92% of patient cases, the automated plans were found to be preferable or comparable to the clinical plans, with 25% rated as preferable. In 67% of cases, the automated plan demonstrated comparable toxicity levels to the clinical plan. Additionally, 47% of cases showed that the automated plans achieve at least a 1% improvement in one of the investigated NTCPs without significantly impacting the others.

This work signifies a more informed strategy for OPT treatment planning, enhancing both speed and automation while reducing planner biases. The methodology offers treatment planners with direct and intuitive information correlating the orientation of the treated eye with toxicity risks. Our investigation considering the direct application of NTCP models demonstrates that eye orientations for OPT could be automatically generated with a comparable or superior treatment suitability for the 92% of investigated patients.