Physics-aware learning and domain-specific loss design in ophthalmology

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Cataracts are the leading cause of blindness globally, necessitating frequent cataract surgery.
• Accurate intraocular lens (IOL) power calculation is crucial for successful refractive outcomes post-surgery.
• Current methods using OCT imaging and machine learning often lack integration of physical principles and sufficient data coverage.

Purpose of the Study:

• To develop a novel optical refraction network, OpticNet, for precise IOL power prediction.
• To integrate physical information and eye optics into a machine learning model for improved accuracy.
• To address challenges of small datasets and domain coverage in machine learning for ophthalmic applications.

Main Methods:

• Developed OpticNet, an unsupervised optical refraction network with a domain-specific loss function.
• Incorporated a differentiable light propagation eye model to backpropogate physical gradients.
• Utilized a transfer learning procedure for unsupervised pre-training and fine-tuning on limited patient data.

Main Results:

• OpticNet outperformed the state-of-the-art on five OCT-image based datasets.
• The proposed method demonstrated improved domain coverage in its predictions.
• Achieved enhanced physical consistency in IOL power calculations.

Conclusions:

• OpticNet offers a significant advancement in predicting IOL power for cataract surgery.
• The integration of physical principles and transfer learning enhances prediction accuracy and reliability.
• This approach holds promise for improving refractive outcomes in patients undergoing cataract surgery.