OPT-former: A transformer-based approach for eye motion tracking in ocular proton therapy.
Authors
Affiliations (4)
Affiliations (4)
- Department of Electronics, Information and Bioengineering, Politecnico di Milano University, Milan, Italy.
- Bioengineering Unit, Clinical Department, National Center for Oncological Hadrontherapy (CNAO), Pavia, Italy.
- Radiotherapy Unit, Clinical Department, National Center for Oncological Hadrontherapy (CNAO), Pavia, Italy.
- Medical Physics Unit, Clinical Department, National Center for Oncological Hadrontherapy (CNAO), Pavia, Italy.
Abstract
Ocular proton therapy (OPT) is the main conservative treatment option for ocular melanoma but requires precise gaze stability monitoring during irradiation. In current clinical workflows at the Centro Nazionale di Adroterapia Oncologica (CNAO) in Pavia, this task is performed manually by trained operators, which can introduce reaction delays and human error. To propose and assess the performance of OPT-former, a transformer-based method for real-time segmentation of the iris and pupil, designed for integration into automated safety interlock systems. The model, inspired by the MedFormer architecture, performs direct three-class segmentation (background, iris, pupil) on 256 × 256 images acquired by CNAO's Eye Tracking System. A dataset of 1683 frames from 173 patients was split into independent training (1349 frames), validation (110 frames), and testing (224 frames) sets. Data augmentation was applied to increase variability, resulting in 3902 training images. Performance was evaluated using Intersection over Union (IoU), Dice coefficient, and Szymkiewicz-Simpson coefficient (SSC), assessing the OPT-former performance and comparing them with a previously proposed 3-step U-Net and manual expert segmentations. Patient-specific fine-tuning was implemented using pre-treatment dry-run session videos. The OPT-former achieved segmentation accuracy comparable to the 3-step U-Net, but with substantially lower inference times (8.62 Hz on GPU vs. 3.26 Hz), enabling near frame-by-frame real-time analysis. Patient-specific fine-tuning significantly improved all metrics for both structures (p < 0.05), with the largest gains for the iris. After fine-tuning, the model outperformed manual annotations from two experts under simulated clinical conditions. The proposed OPT-former achieves high segmentation accuracy, with reduced pipeline complexity and inference time suitable for clinical application. These results endorse its role as a robust support tool in clinical practice, seamlessly integrating into established workflows without impacting the consolidated routine, and enabling automated gaze monitoring and real-time safety interlock activation in OPT.