Deep Learning Vasculature Labeling via Iterative Centerline Refinement

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for classifying patient-specific vasculature require extensive manual work and have limited accuracy due to morphological variations among patients, necessitating a more efficient and accurate automatic detection and labeling system.

Innovation Solution

A deep learning algorithm is trained with a small initial dataset of diagnostic images to identify and label vessels, using centerline extraction and semi-supervised learning to adjust the model based on patient-specific anatomical variations, gradually refining the classification with minimal user input.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If a standard model of vasculature is used for automatic vessel detection and labeling, then the process is automated and manual work is reduced, but the accuracy is limited due to morphological variations among patients

Engineering Contradiction:
Improveautomatic vessel detection and labelingVSAvoiddetection accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The system dynamically adapts the initial standard model by iteratively refining vessel centerlines and adjusting model parameters based on patient-specific imaging data. The model transitions from a static standard representation to a dynamic, patient-specific representation through multiple refinement cycles that incorporate actual anatomical observations.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system modifies key parameters of the standard model including vessel centerline positions, radii, and geometric characteristics to match patient-specific anatomy. By changing these parameters iteratively based on imaging data and deviation analysis, the model achieves high accuracy while maintaining automated operation.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If manual annotation of vessels is performed to improve model accuracy, then detection precision increases, but time consumption and manual effort increase significantly

Engineering Contradiction:
Improvevessel classification accuracyVSAvoidtime for manual work
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary automatic detection and labeling using the initial standard model before any manual intervention. This preliminary action provides a baseline classification that is then automatically refined through iterative processes, minimizing the need for subsequent manual annotation while maintaining high accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system performs self-refinement by automatically detecting deviations between the standard model and patient-specific anatomy, then autonomously adjusting the model parameters and centerlines. This self-service capability reduces reliance on manual annotation while achieving high detection precision.

Inventive Principle:
Principle #25Self-service

3Measurement precision

If a large training dataset is used to train the deep learning algorithm, then model accuracy improves, but the complexity and resource requirements increase

Engineering Contradiction:
Improveclassification accuracyVSAvoidtraining data requirements
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system uses a pre-trained deep learning model with weights initialized from a large training dataset before deployment. This preliminary training action allows the model to achieve high accuracy without requiring access to large datasets during actual operation, reducing complexity and resource requirements at the application stage.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system copies knowledge from a large training dataset into the pre-trained model weights, then applies this copied knowledge to patient-specific cases. This approach transfers the benefits of large-scale training to individual patient analysis without requiring the full dataset to be present during clinical use.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3788631B1Interactive coronary labeling using interventional x-ray images and deep learning
Publication Date: 2025.04.02 KONINKLIJKE PHILIPS NV
  • EP3788631B1 patent drawingFigure 1
  • EP3788631B1 patent drawingFigure 2
  • EP3788631B1 patent drawingFigure 3

AI summary

A method for classifying a vasculature comprises training a training device with an initial model of a vasculature using diagnostic image data representing a geometry for a plurality of vessels of a vessel tree and including a respective vessel labeling for each vessel, providing at least one diagnostic image of a patient's vessel tree and identifying a variation between the vessel tree represented by the initial model and the patient's vessel tree. This variation is checked and labeled in order to improve the trained model. The process may be repeated iteratively until reaching an accurate patient-specific model of the vasculature.