Retinal OCT Layer Mapping via Line-Finding Algorithm

The patent segments the retinal image processing task into distinct stages: initial layer detection using speeded-up robust features (SURF) to identify key anatomical landmarks, followed by separate processing for individual retinal layers. This division allows the computationally intensive CNN to be applied only to critical segmentation points rather than the entire image, reducing overall processing time while maintaining accuracy.

The patent performs preliminary processing by detecting prominent retinal layers and anatomical landmarks using efficient feature detection algorithms before applying CNN-based segmentation. This preliminary action identifies key structures that guide subsequent CNN processing, reducing the computational burden on the neural network and accelerating overall processing while preserving mapping accuracy.

The patent develops a universal CNN-based segmentation approach that can process OCT images from multiple imaging systems using a single trained model. The system achieves this by training on diverse datasets from different OCT platforms and incorporating system-specific parameter adjustments that allow the same network architecture to adapt to various imaging characteristics without requiring complete retraining for each system.

The patent enables adaptation to different OCT imaging systems by modifying network parameters such as learning rate, batch size, and optimization hyperparameters rather than retraining the entire model. This approach allows the same CNN architecture to be efficiently adapted to different systems through parameter tuning rather than full retraining, reducing device complexity while maintaining segmentation accuracy.

The patent extracts and utilizes only the most critical and informative features from training data for CNN segmentation, rather than processing entire images. By identifying and extracting key anatomical landmarks and layer boundaries as primary training targets, the system achieves high segmentation performance with reduced computational resources compared to training on complete retinal images.

The patent applies partial action by focusing CNN training on specific critical regions and layers of the retina that are most important for accurate segmentation, rather than uniformly processing all retinal structures. This selective approach maintains high segmentation performance for clinically relevant layers while reducing the overall computational burden of training and inference.