OCT Retinal and Choroidal Layer Instance Segmentation Using Mask R-CNN.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) provides detailed retinal and choroidal imaging.
• Manual analysis of OCT images is time-consuming, necessitating automated methods.
• Deep learning, particularly encoder-decoder architectures like U-Net, dominates OCT segmentation research.

Purpose of the Study:

• To explore the efficacy of Mask R-CNN, a region proposal-based instance segmentation method, for retinal OCT image segmentation.
• To investigate the impact of hyper-parameter selection on Mask R-CNN performance.
• To compare Mask R-CNN's segmentation accuracy and efficiency against established methods like U-Net.

Main Methods:

• Application of Mask R-CNN for segmenting retinal layers in OCT images.
• Systematic examination of hyper-parameter optimization for the Mask R-CNN model.
• Comparative analysis of segmentation performance metrics (Dice coefficient, boundary errors) and inference time against U-Net.

Main Results:

• Mask R-CNN achieved high Dice coefficients and low segmentation boundary errors, comparable to U-Net.
• The method demonstrated a 2.5-fold reduction in inference time compared to U-Net due to simpler boundary extraction.
• Mask R-CNN successfully segmented seven retinal layers, highlighting its capability for complex OCT analysis.

Conclusions:

• Mask R-CNN is a viable and efficient alternative for automated retinal OCT image segmentation.
• Its simplified boundary extraction offers significant advantages in reducing processing time for clinical applications.
• Further research into hyper-parameter optimization can enhance its utility in ophthalmological diagnostics.