Blood Vessel Segmentation Using Keypoint Distance Regression

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current convolutional neural networks (CNNs) for image segmentation are limited to pixel or voxel accuracy, introducing quantization errors and failing to account for structures without holes or disconnected components, necessitating models that can achieve sub-pixel or sub-voxel accurate segmentations.

Innovation Solution

A method involving training a model using keypoints and computed distances to predict segmentation boundaries with sub-pixel or sub-voxel accuracy, utilizing a training phase to optimize parameters and a testing phase to segment structures with continuous regression values.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a CNN is trained to predict class labels for each image element (pixel or voxel), then the segmentation process is automated and can be applied to medical images, but the segmentation boundary accuracy is limited to the level of an image element (pixel or voxel) and cannot achieve sub-pixel or sub-voxel accuracy

Engineering Contradiction:
Improvesegmentation boundary accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent transforms the segmentation problem from predicting class labels at discrete pixel/voxel locations to predicting continuous distance values from keypoints to boundaries. This dimensional change from discrete grid-based predictions to continuous distance field predictions enables sub-pixel accuracy by operating in a continuous mathematical space rather than being constrained to discrete image element boundaries.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The patent introduces keypoints as intermediary elements that serve as reference points for distance computation. Instead of directly predicting boundaries at every pixel location, the model predicts distances from selected keypoints to the actual boundaries, using these distance fields as intermediaries to reconstruct accurate sub-pixel boundary locations through level set methods.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If standard CNN assumptions are used without integrating structural constraints, then the model is simpler to implement, but the predicted segmentation labels may contain spurious components and holes in the segmented objects

Engineering Contradiction:
Improvesegmentation correctnessVSAvoidmodel complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent incorporates structural constraints (such as connectivity and hole-free assumptions) directly into the model architecture and loss function design during the training phase. By preliminarily embedding these anatomical priors into the distance prediction framework and level set evolution process, the model inherently produces topologically correct segmentations without requiring post-processing corrections.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes the fundamental prediction parameter from discrete class labels to continuous distance values. This parameter transformation allows the model to enforce structural constraints through the mathematical properties of distance fields and level set evolution, ensuring that predicted boundaries naturally satisfy connectivity and topological requirements while maintaining model tractability.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4068203B1Systems and methods for anatomic structure segmentation in image analysis
Publication Date: 2026.04.29 HEARTFLOW INC
  • EP4068203B1 patent drawingFigure 1
  • EP4068203B1 patent drawingFigure 2A~2B
  • EP4068203B1 patent drawingFigure 3A~3B

AI summary

Systems and methods are disclosed for anatomic structure segmentation in image analysis, using a computer system. One method includes: receiving (311) image data representing a blood vessel; transforming (312) the image into a curvilinear planar representation and defining a centerline of the blood vessel, including extracting a set of frames along the centerline; defining (313) points on the centerline as keypoints; defining (314) a mapping from the image coordinates to a Euclidean space, including, for each frame, determining (315) curvilinear planar representation intensity values in a set of angular directions around a keypoint; based on the intensity values; predicting (316) the segmentation boundary of the patient's blood vessel using a trained convolutional neural network, said CNN having been trained to predict the distance from a given keypoint to the blood vessel lumen boundary; and outputting (317) a predicted anatomic model and/or a complete segmentation of the blood vessel.