DPLSeg: unsupervised general segmentation model for retinal images across multiple OCT devices.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of retinal diseases like myopic and diabetic retinopathy is crucial.
• Accurate retinal-layer segmentation in optical coherence tomography (OCT) images is essential for diagnosis.
• Current deep learning models struggle with domain shifts and high annotation costs.

Purpose of the Study:

• To develop an unsupervised deep learning model for precise retinal-layer segmentation.
• To improve model generalization across different OCT devices and reduce annotation dependency.
• To enhance early detection of retinal diseases through improved OCT image analysis.

Main Methods:

• Proposed dual-level pseudo-label learning for segmentation (DPLSeg).
• Incorporated a hierarchical transformer encoder for enhanced feature representation and domain adaptability.
• Utilized an unsupervised learning strategy to minimize annotation requirements.

Main Results:

• DPLSeg achieved a mean intersection over union (mIoU) of 79.9% on 850 OCT images from three devices.
• Outperformed existing models like DeepLab (75.2% mIoU) and DAFormer.
• Reduced annotation needs by 80% compared to supervised methods.

Conclusions:

• DPLSeg offers a robust and accurate solution for retinal-layer segmentation in OCT images.
• The model demonstrates strong domain adaptability, addressing challenges posed by multi-device data.
• DPLSeg presents a scalable and cost-effective tool for clinical diagnostics and early retinal disease detection.