OPT-former: A transformer-based approach for eye motion tracking in ocular proton therapy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Ocular proton therapy (OPT) is crucial for ocular melanoma treatment.
• Precise gaze stability monitoring is essential during OPT irradiation.
• Current manual monitoring by operators introduces delays and potential errors.

Purpose of the Study:

• To develop and evaluate OPT-former, a transformer-based method for real-time iris and pupil segmentation.
• To integrate OPT-former into automated safety interlock systems for enhanced precision.
• To improve the safety and efficiency of ocular proton therapy.

Main Methods:

• A transformer-based model (OPT-former) was developed for three-class segmentation (background, iris, pupil).
• The model was trained and tested on a dataset of 1683 frames from 173 patients.
• Performance was evaluated using IoU, Dice, and SSC, with comparisons to a U-Net model and manual segmentations.
• Patient-specific fine-tuning was implemented using pre-treatment dry-run videos.

Main Results:

• OPT-former achieved segmentation accuracy comparable to a 3-step U-Net.
• Inference times were significantly lower (8.62 Hz vs. 3.26 Hz), enabling near real-time analysis.
• Patient-specific fine-tuning substantially improved segmentation metrics for iris and pupil.
• The fine-tuned model surpassed manual expert annotations in simulated clinical conditions.

Conclusions:

• OPT-former offers high segmentation accuracy with reduced complexity and faster inference times for clinical use.
• The model can be seamlessly integrated into existing OPT workflows.
• It enables automated gaze monitoring and real-time safety interlock activation, enhancing clinical practice.