PViTA: Attention-Enhanced Vision Transformer Approach for Analyzing Pupil Dynamics in Asthenopia

The increased usage of digital devices and prolonged close-up tasks leads to Asthenopia, which causes long-term visual impairments, making early diagnosis essential. Analysis of pupil response to light provides valuable insights as its size varies during such tasks. This work focuses on quantifying pupillary size with varying light intensities using the proposed model "PViTA" (Pupil-Based Vision Transformer with enhanced Attention). The customized dataset is formulated with eye images captured using a digital camera under different lighting conditions. They are subjected to cross-domain generalization for segmenting the pupil region using PViTA. This is carried out using a publicly available MOBIUS eye dataset, containing pupil of different sizes, shapes, and occlusions, which are utilized for training the model. The model shows an accuracy of 0.9816 with a dice value of 0.7015 during training. The trained model is tested on the customized dataset, and the segmented mask is utilized to analyze pupil changes during different light intensities by measuring the diameter. The findings demonstrate that the pupil constricts with brighter intensities and commences the adaptation process after crossing a threshold interval. Furthermore, diameter variation in the ground truth and predicted mask is analyzed, and it shows a similar pattern where the pupil size decreases from 65 to 40 pixel units from the minimum to the maximum intensity level. Thus, the proposed framework can be effectively utilized to capture the pupil dynamics during different intensity ranges.