Weakly supervised serous retinal detachment segmentation in SD-OCT images by two-stage learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Automated segmentation of retinal diseases in SD-OCT images is crucial for quantitative assessment.
• Deep convolutional neural networks (CNNs) excel at image segmentation but require extensive pixel-wise annotations.
• Acquiring high-quality annotations for retinal diseases is costly and labor-intensive.

Purpose of the Study:

• To develop a weakly supervised learning architecture for detecting and segmenting central serous chorioretinopathy (CSC) retinal detachment.
• To reduce the reliance on expensive pixel-level annotations in SD-OCT image analysis.
• To enable efficient and accurate quantitative assessment of CSC.

Main Methods:

• A two-stage weakly supervised learning architecture was proposed.
• Stage 1: Located-CNN for lesion detection and highlighting, refined with level set methods for pseudo pixel-level labels.
• Stage 2: Active-contour loss function integrated into deep networks for precise lesion segmentation.

Main Results:

• The proposed method achieved competitive performance compared to models with stronger supervision.
• Demonstrated superior performance over existing models trained with different supervision levels.
• Successfully segmented CSC retinal detachment using only image-level annotations.

Conclusions:

• This work presents the first weakly supervised learning approach for CSC segmentation in SD-OCT images.
• The method significantly minimizes the need for manual annotation, reducing effort and cost.
• The proposed architecture offers a viable and efficient alternative for clinical quantitative assessment of CSC.