Medical Image Landmark Extraction With Relationship-Based Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for extracting anatomical landmarks from medical images using machine learning often result in erroneous detection of anatomical landmarks due to the lack of consideration for the relationships between these landmarks, leading to incomplete or inaccurate landmark extraction.
Innovation Solution
A medical image processing apparatus that identifies landmark groups based on anatomical tissue relationships and corrects landmark positions using continuity and pixel value appropriateness indices to enhance extraction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning is used to extract anatomical landmarks from medical images, then extraction speed is improved, but detection accuracy deteriorates due to erroneous detection
Solution Approach 1:
The patent implements a feedback mechanism where extracted landmarks are evaluated against continuity constraints and pixel value appropriateness indices. Landmarks that violate these constraints are identified as erroneous and corrected through iterative optimization, allowing the system to maintain high extraction speed while improving detection accuracy through self-correction.
Solution Approach 2:
The patent replaces pure data-driven machine learning extraction with a hybrid approach that incorporates physics-based constraints (continuity of anatomical structures and pixel value gradients). This substitution of mechanical/physical reasoning for purely statistical learning enables the system to correct erroneous detections while maintaining processing efficiency.
2Loss of time
If conventional machine learning extraction is used, then processing time is reduced, but landmark position accuracy deteriorates
Solution Approach 1:
The patent performs preliminary extraction using machine learning to obtain initial landmark positions quickly, then applies continuity constraints and pixel value checks as preliminary validation steps before final output. This preliminary action filters out obviously erroneous detections without requiring complete re-extraction, thus minimizing time loss while improving position accuracy.
Solution Approach 2:
The patent changes the evaluation parameters from purely statistical learning metrics to include continuity constraints and pixel value appropriateness indices. By modifying the validation parameters to incorporate anatomical plausibility checks, the system can correct position errors without reverting to slow re-extraction processes.
3Device complexity
If anatomical landmark extraction is performed without considering relationships between landmarks, then extraction simplicity is maintained, but extraction completeness deteriorates
Solution Approach 1:
The patent segments the extraction process into independent machine learning extraction of individual landmarks, followed by separate continuity constraint validation and pixel value appropriateness checking. This segmentation allows the system to maintain simple individual extraction while adding relationship-based completeness checks as separate, modular steps that do not significantly increase overall complexity.
Solution Approach 2:
The patent introduces continuity constraints and pixel value appropriateness indices as intermediary validation mechanisms between the simple extraction process and the final landmark set. These intermediaries act as filters that ensure relationship consistency without requiring the extraction process itself to become complex, thus maintaining simplicity while improving completeness.
Data Source
AI summary
A medical image processing apparatus includes processing circuitry. The processing circuitry extracts, from a medical image, anatomical landmarks that represent feature points of anatomical tissue that is included in the medical image, identifies a landmark group to which the extracted anatomical landmarks belong from among landmark groups in each of which a plurality of anatomical landmarks that are groped based on the anatomical tissue and relationship information that defines a physical relationship among the anatomical landmarks are associated, and corrects positions of the extracted anatomical landmarks on the medical image based on the relationship information on the identified landmark group.