Adaptive Tensor-Based Feature Extraction for Pupil Segmentation in Cataract Surgery.

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging

Background:

• Cataract surgery is the sole treatment for visually significant cataracts, a leading cause of preventable blindness.
• Stable pupil dilation is crucial for successful cataract surgery.
• Automated pupil segmentation aids in identifying risks for pupillary instability during surgery.

Purpose of the Study:

• To introduce a novel adaptive wavelet tensor feature extraction (AWTFE) method to improve deep learning-based pupil segmentation accuracy in cataract surgery.
• To address challenges like variable illumination, instrument obstruction, and lens hydration affecting pupil recognition.

Main Methods:

• Constructing a third-order tensor to represent spatial, color, and wavelet subband correlations.
• Employing higher-order singular value decomposition for adaptive redundant information elimination and pupil feature estimation.
• Evaluating AWTFE with deep learning models on the BigCat and CaDIS datasets.

Main Results:

• AWTFE significantly improved segmentation performance by up to 3.31% on the CaDIS dataset and 2.26% on the BigCat dataset.
• The method achieved statistically significant improvements (P < 1.29 × 10^-10) across models, reaching Dice coefficients of 94.74% (BigCat) and 96.71% (CaDIS).
• AWTFE enhanced performance by up to 2.87% during challenging surgical phases and outperformed other feature extraction techniques.

Conclusions:

• The AWTFE method effectively extracts relevant pupil features, enhancing deep learning model accuracy for pupil segmentation.
• This approach offers a robust solution for improving automated pupil recognition in the complex environment of cataract surgery.
• AWTFE contributes to safer cataract surgery by enabling more reliable pre-operative risk detection for pupillary instability.