Deep Learning-Based Automated Pupil Segmentation In Cataract Surgery Images Using U-Net Architecture

Accurate pupil segmentation during cataract surgery is a critical prerequisite for computer-assisted surgical guidance and intraoperative analytics. Manual delineation of pupil boundaries is time-consuming, subjective, and impractical for real-time applications. Here we present an end-to-end deep learning framework for fully automated pupil segmentation from intraoperative cataract surgery images.

Single tertiary referral eye centre

We implemented a U-Net architecture with a four-level encoder (64–128–256–512 channels), 1,024-channel bottleneck, and symmetric decoder with skip connections (31,043,521 parameters). The Cataract-1K dataset (2,256 annotated images) was partitioned into training (70%; n=1,579), validation (15%; n=338), and test (15%; n=339) sets. Images were resized to 512×512 pixels and normalised using ImageNet statistics. Data augmentation comprised random horizontal flipping (p=0.5). The model was optimised using AdamW (learning rate=1×10⁻⁴) with combined Dice–binary cross-entropy loss. A ReduceLROnPlateau scheduler (patience=5) adjusted learning rates. Training proceeded for 15 epochs with early stopping (patience=15).

On the held-out test set of 339 images, the model achieved a mean Dice coefficient of 0.9805 ± 0.0553 and a mean IoU of 0.9648 ± 0.0586, with a precision of 0.9825 ± 0.0550 and a recall of 0.9788 ± 0.0587. Pixel-level evaluation yielded an ROC-AUC of 0.9995 and a PR-AUC of 0.9979, while overall pixel accuracy reached 0.9973 ± 0.0030. The best validation Dice score (0.9800) was obtained at epoch 14, with validation performance exceeding 0.97 by epoch 5, indicating rapid convergence and stable generalization.

These findings indicate that a standard U-Net architecture, without pre-trained encoders or complex augmentation strategies, can achieve highly accurate pupil segmentation in cataract surgery images. The consistently high Dice and IoU scores, together with near-unity ROC-AUC, demonstrate robust boundary delineation across diverse surgical conditions. Minor performance degradation was observed in cases with severe instrument occlusion or atypical illumination. Future work will focus on incorporating attention mechanisms, transfer learning, and multi-class segmentation of additional ocular structures to further enhance clinical applicability.