Localization of Polypoidal Choroidal Vasculopathy in Fluorescein Angiography Using Semisupervised Deep Learning With...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Polypoidal choroidal vasculopathy (PCV) is a significant cause of vision impairment.
• Accurate segmentation of PCV lesions in fluorescein angiography (FA) is crucial for diagnosis and treatment monitoring.
• Traditional methods often require pixel-level annotations or additional imaging modalities like indocyanine green angiography (ICGA).

Purpose of the Study:

• To develop a cost-effective, semisupervised deep learning model for segmenting PCV in FA images.
• To leverage a large dataset of unlabeled FA images to improve segmentation accuracy.
• To reduce the reliance on pixel-level annotations and ICGA for PCV analysis.

Main Methods:

• A variant of the mean teacher (MT) model, a semisupervised convolutional neural network, was developed.
• The model utilizes both labeled and unlabeled FA images for training.
• Two identical networks (student and teacher) were employed for aggressive learning and consistency regulation.

Main Results:

• The MT model achieved Dice similarity coefficients of 0.577 (validation) and 0.518 (testing).
• Hausdorff distances were 37.28 pixels (validation) and 43.05 pixels (testing).
• Even with halved labeled data, the MT model outperformed AG-PCVNet, a specialized recurrent convolutional neural network.

Conclusions:

• Semisupervised deep learning effectively segments PCV lesions using abundant unlabeled FA images.
• The model eliminates the need for ICGA in training for PCV monitoring.
• Improved accuracy was observed in cases with hyperfluorescence compared to fully supervised models.