KEY POINTS
- The study developed a transformer-based model for automated iris and pupil segmentation during ocular proton therapy, where patients must maintain a prescribed gaze while trained operators currently monitor eye position visually during beam delivery.
- The dataset contained 1,683 video frames from 173 ocular melanoma patients treated at CNAO. Independent sets comprised 1,349 training frames, 110 validation frames and 224 test frames; augmentation expanded the training dataset to 3,902 images.
- On the independent test set, OPT-former achieved median pupil IoU 0.902, Dice 0.949 and SSC 0.960, while iris performance was IoU 0.745, Dice 0.854 and SSC 0.920. Segmentation accuracy was statistically comparable with the previous three-step U-Net pipeline.
- The major advantage was speed and workflow simplicity. OPT-former performs background, iris and pupil segmentation in a single network rather than sequential ROI, iris and pupil models. GPU processing required a median 116 ms per frame, versus 310 ms with the three-step U-Net, corresponding to 8.62 versus 3.26 frames/s (p<10⁻⁶).
- Patient-specific fine-tuning using pretreatment dry-run video frames improved IoU and Dice significantly for both iris and pupil in all six tested patients. The largest improvement occurred in a difficult case where iris IoU increased from approximately 0.64 to 0.94 and Dice from 0.78 to 0.97.
- After fine-tuning, pupil IoU across individual patients ranged from 0.95 to 0.99, while iris IoU ranged from 0.83 to 0.94. In a separate 178-frame clinical-like assessment, the fine-tuned algorithm significantly outperformed two sets of manual operator annotations across the evaluated segmentation metrics.
- The model was designed as a building block for an automated beam safety interlock, but the study did not measure the complete camera-to-beam-interruption latency or prospectively test automatic beam stopping during treatment. The results therefore validate segmentation speed and accuracy, not the complete clinical safety system.
CLINICAL TAKEAWAY
This addresses a very practical ocular-proton problem: replacing continuous subjective gaze monitoring with fast quantitative tracking capable of feeding an automated interlock. The algorithm is fast enough for the existing 6–12-Hz video stream, but prospective validation of actual movement detection and beam interruption is the critical next step.