Nested Two-Class Classifiers for Anatomical Landmark Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for locating anatomical landmarks in medical image data using multi-class classifiers face inaccuracies due to the challenge of representing diverse background regions with a small number of training points, leading to imprecise landmark detection.

Innovation Solution

The use of nested two-class classifiers, where a first classifier distinguishes between a larger foreground and background region, and a second classifier is trained for a smaller foreground region within the first, allowing for a balanced number of background and foreground training points, thereby improving accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a multi-class classifier is used to identify multiple anatomical landmarks simultaneously, then the efficiency of landmark identification is improved, but the accuracy deteriorates due to the inability to adequately represent diverse background regions with a small number of training points

Engineering Contradiction:

Improveefficiency of landmark identificationVSAvoidaccuracy of landmark location

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent divides the single multi-class classification task into multiple sequential two-class classification stages. Each stage focuses on distinguishing a specific foreground region from its corresponding background region, thereby simplifying the classification problem and improving accuracy while maintaining efficiency through automated processing

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If foreground regions are restricted to a small region around each landmark to achieve better accuracy, then the precision of landmark location is improved, but the number of training points decreases, requiring a similarly limited number of background training points which must represent a huge variation in background voxels

Engineering Contradiction:

Improveprecision of landmark locationVSAvoidcomplexity of representing background variation

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the classification process into multiple stages, where each stage handles a specific foreground-background distinction. This allows the system to use small, focused foreground regions for precision while managing background representation in a controlled manner through sequential classification rather than attempting to represent all background variation simultaneously

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a temporal dimension by using sequential classification stages rather than a single simultaneous classification. This allows the system to progressively refine landmark location through multiple passes, effectively adding a time dimension to the classification process that resolves the contradiction between precision and background representation complexity

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Data Source

PatentUS9235781B2Method of, and apparatus for, landmark location

Publication Date: 2016.01.12 TOSHIBA MEDICAL SYST CORP

AI summary

An apparatus for locating a landmark in a set of image data comprises a landmark location unit that is configured, for each of a plurality of image data items, to obtain from a first two-class classifier a first classification of the image data item as foreground or background, to obtain from a second two-class classifier a second classification of the image data item as foreground or background, and to combine the first classification and the second classification to obtain a combined classification, and wherein the landmark location unit is further configured to use the combined classifications for the plurality of image data items to determine a location for the landmark.