Localization of Polypoidal Choroidal Vasculopathy in Fluorescein Angiography Using Semisupervised Deep Learning With Labeled and Unlabeled Images.

Purpose: To determine the extent of polypoidal choroidal vasculopathy (PCV) in fluorescein angiography (FA) using cost-effective, semisupervised deep learning to leverage abundance of FA images with PCV condition without pixel-level annotation. Methods: We develop a variant of the mean teacher (MT) model, which is a semisupervised convolutional neural network trained with both labeled and unlabeled data. The model is composed of two identical networks: a student network, learning aggressively if given the labeled images, and is regulated by a teacher network for consistency, if given unlabeled data. Results: Our MT model is validated with a dataset containing images from the EVEREST-I study and 3 medical centers in Asia, for a total of 146 labeled eyes and 415 unlabeled eyes. It achieves Dice similarity coefficients of 0.577 and 0.518 for validation and testing, respectively, and 37.28 and 43.05 pixels, respectively, for Hausdorff distance. If the labeled data are further reduced by one-half, with the support of the unlabeled data, the MT model still outperforms a recurrent convolutional neural network model designed for PCV segmentation using FA sequences, named AG-PCVNet, trained with the full set in both Dice similarity coefficient and Hausdorff distance. Conclusions: The proposed semisupervised model demonstrated the efficacy of deep learning for PCV lesion segmentation with abundance of unlabeled FA images without matching indocyanine green angiography images. Significant improvement in accuracy is observed in cases with hyperfluorescence, compared with a fully supervised model. Translational Relevance: The developed deep learning system eliminates the need for indocyanine green angiography in training an effective model for monitoring polypoidal choroidal vasculopathy treatment.