Lightweight Learning-Based Automatic Segmentation of Subretinal Blebs on Microscope-Integrated Optical Coherence...

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Subretinal injections are vital for treating eye diseases, but precise therapeutic dosing is challenging due to leakage.
• Current methods lack the ability to measure the volume of therapeutics successfully delivered to the subretinal space during surgery.

Purpose of the Study:

• To develop and validate the first automatic method for quantifying subretinal bleb volumes.
• To enable accurate assessment of therapeutic delivery in subretinal injections.

Main Methods:

• An ex vivo porcine eye model was used, with subretinal injections of Ringer's lactate solution.
• Microscope-integrated optical coherence tomography provided 3D visualization of subretinal blebs.
• A novel deep learning algorithm was developed to segment bleb boundaries, with cross-validation and ensemble-classifier strategies employed.

Main Results:

• The deep learning algorithm achieved high accuracy in segmenting subretinal bleb boundaries.
• Achieved F1 scores of 93.86 ± 1.17% for entry blebs and 96.90 ± 0.59% for full blebs on independent test data.
• Demonstrated superior performance compared to four other state-of-the-art deep learning segmentation methods.

Conclusions:

• The proposed algorithm accurately quantifies subretinal bleb volumes in porcine eyes with robust performance and real-time speed.
• This represents a significant advancement for computer-guided therapeutic delivery in the subretinal space.
• Enables precise measurement of injected fluid volumes, crucial for optimizing ocular drug delivery.