Semi-Supervised Dual Stream Segmentation Network for Fundus Lesion Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of retinal images is crucial for diagnosing retinopathy and systemic diseases.
• Complex retinal structures and lesions pose challenges for current segmentation methods.
• Existing techniques struggle with precise identification of specific lesions like capillary non-perfusion and choroidal neovascularization.

Purpose of the Study:

• To propose a novel dual stream segmentation network within a conditional generative adversarial network (cGAN) framework.
• To enhance the accuracy of retinal lesion segmentation in fundus fluorescein angiography (fFIA) and optical coherence tomography (OCT) images.
• To leverage both labeled and unlabeled data for improved segmentation performance through semi-supervised adversarial training.

Main Methods:

• A dual stream encoder architecture was developed to extract comprehensive feature information from retinal images.
• A multiple level fuse block was designed to effectively decode features from parallel encoders.
• Semi-supervised adversarial training, incorporating a dual stream Bayesian segmentation network and an annotation discriminator, was employed.

Main Results:

• The proposed dual stream network demonstrated superior segmentation accuracy for retinal capillary non-perfusion regions and choroidal neovascularization.
• Cross-validation on 384 fFIA and 1040 OCT images confirmed the method's effectiveness.
• The approach outperformed state-of-the-art methods in segmenting complex retinal pathologies.

Conclusions:

• The proposed dual stream segmentation network offers a significant advancement in automated retinal image analysis.
• This method holds promise for improving the accuracy and efficiency of diagnosing various eye conditions.
• The integration of cGANs and semi-supervised learning provides a robust strategy for medical image segmentation.