CT Lesion Segmentation Decision Rule
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CAD systems for analyzing CT scans face challenges in accurately estimating the shape of lesions like lung nodules and colon polyps, often incorrectly attaching them to anatomic structures during segmentation, leading to poor feature extraction and incorrect classification.
Innovation Solution
A decision rule or classifier is used to examine CAD-detected regions of interest in each CT slice pair to determine if they belong to the same object, refining segmentation boundaries before feature calculation by combining or splitting adjacent slices based on mutual overlap features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If simple connectiveness rule is used to form objects from voxels, then object formation is computationally simple, but adjacent slices are incorrectly joined when voxels are far apart in Z dimension
Solution Approach 1:
The patent changes the parameter from simple spatial connectiveness to a decision rule based on multiple features including mutual overlap, intensity correlation, and shape similarity. This transforms the object formation process from a geometric operation to a multi-parameter classification problem, resolving the contradiction between computational simplicity and segmentation accuracy.
Solution Approach 2:
The patent introduces a decision rule or classifier as an intermediary between voxel detection and object formation. This intermediary evaluates multiple features and makes intelligent decisions about whether adjacent slices belong to the same object, preventing incorrect joins while maintaining systematic processing.
2Manufacturing precision
If mathematical morphological opening is used to break connections, then some attachments are removed, but the operation is insufficient when region of interest and anatomic structure have high overlap
Solution Approach 1:
The patent moves from a single-parameter morphological operation to a multi-parameter decision rule that evaluates mutual overlap, intensity correlation, and shape similarity. This allows the system to handle cases of high overlap by considering multiple features simultaneously rather than relying on a single morphological operation.
Solution Approach 2:
The patent segments the segmentation process into distinct stages: initial detection, refined segmentation using decision rules, and final object formation. This multi-stage approach allows mathematical morphological opening to be used in conjunction with more sophisticated decision rules, combining the benefits of both methods.
3Productivity
If detector ROIs are used as is, then detection speed is maintained, but the ROIs underestimate the complete extent of objects leading to poor feature separability
Solution Approach 1:
The patent performs refined segmentation as a preliminary action before feature calculation and classification. By expanding and refining the detector ROIs before feature extraction, the system ensures that subsequent measurements are based on accurate object extents, improving feature separability without sacrificing the initial detection speed.
Solution Approach 2:
The patent divides the processing into distinct segmentation stages: initial detector ROI generation, refined segmentation with decision rules, and final object formation. This segmentation allows the system to maintain fast initial detection while applying more computationally intensive refinement only where needed.
Data Source
AI summary
A decision rule is used that examines the computer-aided detected (CAD) regions of interest in a computed tomography (CT) slice pair taken from volumetric medical CT scan to determine whether the detected regions of interest are part of the same object is disclosed. Segmentation is performed after initially detecting a region of interest but before calculating features in order to refine the boundaries of the detected regions of interest. Segmentation occurs in the two-dimensional slices by segmenting the region of interest on each slice. Adjacent slices are examined to determine if adjacent objects are actually part of the same structure. If they are not, the objects are split apart. In this way, three-dimensional objects are formed from two-dimensional segmentations.