Multi-atlas Segmentation via Downsampled Label Fusion

The patent segments the computational process into two distinct phases: (1) a fast coarse segmentation phase using simplified label propagation to capture major anatomical structures, and (2) a refined correction phase using deformable registration only in critical regions. This segmentation allows the system to achieve high overall accuracy while avoiding the prohibitive computational cost of applying full deformable registration throughout the entire volume.

The patent applies deformable registration selectively rather than universally - specifically, it applies the computationally intensive correction only in regions where segmentation uncertainty is high or anatomical variability is expected. This partial application of the refined method maintains accuracy where needed while dramatically reducing overall computational burden compared to exhaustive application.

The patent divides memory usage into two stages: first, loading and processing a limited set of atlas images and their segmentations for coarse label propagation; second, selectively loading high-resolution deformable registration data only for specific regions of interest. This segmentation of memory requirements allows the system to achieve high accuracy without requiring excessive memory resources throughout the entire processing pipeline.

The patent extracts and processes only the essential components needed for accurate segmentation - specifically, it extracts key anatomical landmarks and boundary information from atlas images, then uses only these extracted features for the deformable registration phase. This extraction approach reduces memory requirements by eliminating the need to store and process entire high-resolution atlas volumes.

The patent performs preliminary coarse segmentation using fast label propagation methods to establish an initial anatomical framework and identify regions of interest. This preliminary action creates a foundation that guides subsequent refined processing, ensuring that the faster method accomplishes useful work before the slower, more accurate method is applied to critical regions.

The patent introduces an intermediary correction step that bridges the fast coarse segmentation and the final accurate result. This intermediary phase uses the outputs from fast label propagation (such as initial segmentation masks and registration transforms) as inputs to a more refined deformable registration process, thereby combining the speed of fast methods with the accuracy of slow methods.