SD-RetinaNet: Topologically Constrained Semi-Supervised Retinal Lesion and Layer Segmentation in OCT

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is crucial for diagnosing and monitoring retinal diseases.
• Accurate segmentation of retinal layers and lesions is vital for patient care.
• Existing semi-supervised methods struggle with anatomical plausibility and topological correctness.

Purpose of the Study:

• To develop a novel semi-supervised model for accurate retinal biomarker segmentation.
• To enforce anatomical correctness and topological guarantees in segmentation.
• To improve the modeling of interactions between retinal layers and lesions.

Main Methods:

• A fully differentiable biomarker topology engine was introduced to enforce anatomical constraints.
• Joint learning with bidirectional influence between layers and lesions was enabled.
• A disentangled representation separating spatial and style factors was learned.

Main Results:

• The model achieved state-of-the-art performance in both lesion and layer segmentation.
• It demonstrated improved realism in layer segmentation and accurate lesion localization.
• The model generalized layer segmentation to pathological cases using partially annotated data.

Conclusions:

• The proposed model offers accurate, robust, and trustworthy retinal biomarker segmentation.
• Incorporating anatomical constraints into semi-supervised learning is highly effective.
• This approach has significant potential for improving OCT-based disease diagnosis and monitoring.