Segment Anything in Optical Coherence Tomography: SAM 2 for Volumetric Segmentation of Retinal Biomarkers

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is crucial for non-invasive retinal imaging and disease detection.
• Artificial intelligence (AI), especially deep learning (DL), is increasingly applied in medical imaging.
• Accurate segmentation of biomarkers in OCT scans is vital for diagnosing retinal diseases like age-related macular degeneration (AMD) and diabetic macular edema (DME).

Purpose of the Study:

• To evaluate the performance of new AI models, SAM 2 and MedSAM 2, for segmenting biomarkers in 3D OCT volumes.
• To compare the segmentation accuracy of SAM 2 and MedSAM 2 against the traditional U-Net model.
• To assess the utility of these AI models in enhancing the quality of retinal disease diagnostics.

Main Methods:

• Utilized SAM 2 and MedSAM 2 for segmenting OCT volumes from two open-source datasets (OIMHS and AROI).
• Employed the U-Net model as a baseline for performance comparison.
• Quantified segmentation performance using the Dice score for specific biomarkers: macular holes (MH), intraretinal cysts (IRC), intraretinal fluid (IRF), and pigment epithelial detachment (PED).

Main Results:

• SAM 2 and MedSAM 2 achieved high Dice scores, indicating excellent segmentation accuracy.
• Achieved an overall Dice score of 0.913 for MH and 0.902 for IRC on the OIMHS dataset.
• Obtained Dice scores of 0.888 for IRF and 0.909 for PED on the AROI dataset.

Conclusions:

• SAM 2 and MedSAM 2 demonstrate superior or comparable performance to U-Net for OCT biomarker segmentation.
• These advanced AI models offer a more efficient and accurate alternative to manual segmentation for 3D OCT scans.
• The findings suggest significant potential for AI-driven segmentation in improving the early detection and monitoring of retinal diseases.