Anatomy-Aligned Image Patches for Accurate Medical Grading

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for grading conditions of anatomical structures in medical images, such as cartilage damage, often miss critical medical information due to insufficient determination of findings, necessitating improved methods for generating training data for neural-network based classifiers.

Innovation Solution

A method for generating anatomy-aligned image patches by analyzing medical images to determine regions of interest and using landmarks, which are encoded into a model-based segmentation to create aligned image patches for training classifiers, allowing for increased training data and improved accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional segmentation methods are used to detect findings in medical images, then the process is simple and fast, but critical medical information is missed due to insufficient determination of findings

Engineering Contradiction:
Improvefinding detection accuracyVSAvoidanalysis method complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the anatomical structure into multiple regions of interest (ROIs) based on landmark-based segmentation. Each ROI is then individually analyzed by the classifier to detect findings. This segmentation approach enables more precise local analysis without requiring overly complex global analysis methods, resolving the contradiction between detection accuracy and method complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary landmark identification and ROI segmentation before the actual finding detection. By pre-defining the regions of interest based on anatomical landmarks, the system prepares the data structure in advance, which simplifies the subsequent classification process while improving finding detection accuracy through focused regional analysis.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If more image data is collected to improve classifier training, then grading accuracy improves, but data collection time and storage requirements increase

Engineering Contradiction:
Improvegrading accuracyVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments anatomical structures into standardized regions of interest using landmark-based methods. This segmentation enables the classifier to be trained on focused regional data rather than requiring extensive full-image datasets. The segmented ROIs provide consistent, anatomically-aligned training samples that improve grading accuracy while reducing the overall data collection burden.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different analysis approaches to different regions of interest based on their anatomical characteristics. Each ROI is extracted and analyzed with appropriate classification parameters, allowing the system to achieve high grading accuracy through specialized local analysis rather than requiring uniformly high-quality data across entire images.

Inventive Principle:
Principle #3Local quality

3Manufacturing precision

If anatomical landmarks are precisely determined to define regions of interest, then image patch alignment improves, but processing complexity increases

Engineering Contradiction:
Improveimage patch alignment precisionVSAvoidlandmark determination complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent performs landmark identification and ROI segmentation as preliminary steps before the main classification task. By establishing the anatomical framework in advance, the system achieves precise image patch alignment without adding complexity to the core finding detection process. The preliminary landmark-based segmentation creates a structured basis for subsequent analysis.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses anatomical landmarks as intermediary reference points to define regions of interest. These landmarks serve as mediators between the raw medical image and the final finding detection, providing a standardized anatomical framework that simplifies the alignment process while maintaining high precision through biologically meaningful reference points.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4604056A1Anatomy-aligned image patches for grading a condition of an anatomical structure
Publication Date: 2025.08.20 KONINKLIJKE PHILIPS NV
  • EP4604056A1 patent drawingFigure 1
  • EP4604056A1 patent drawingFigure 2
  • EP4604056A1 patent drawingFigure 3A~3D

AI summary

The present invention relates to a (computer-implemented) method for generating training data for training a classifier. The method comprises receiving a medical image comprising an anatomical structure of a subject, analyzing the medical image with respect to a positioning of the anatomical structure in the medical image, and determining based on a result of the analyzation at least a first region of interest of the anatomical structure and a second region of interest of the anatomical structure. The method comprises further generating at least a first image patch corresponding to the first region of interest of the anatomical structure and a second image patch corresponding to the second region of interest of the anatomical structure, and providing the first image patch and the second image patch as training data for training the classifier.