3D Vessel Depth Reconstruction from 2D Medical Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional centerline tracing methods in medical images fail to distinguish between bifurcated and overlapping vessels, leading to incorrect interpretations and false path tracing in 2D angiographic images.
Innovation Solution
A multi-channel depth image is generated using a series of trained machine learning networks, including an image-to-image network for branch overlap prediction, a fully convolutional neural network for branch orientation analysis, and another image-to-image network for depth reconstruction, to accurately represent vessel structures and differentiate between bifurcations and overlaps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional centerline tracing methods are used, then the process is simple and fast, but the method cannot distinguish between bifurcated vessels and overlapping vessels, leading to incorrect interpretations
Solution Approach 1:
The patent transforms 2D medical images into 3D depth representations by adding depth information as a new dimension. Multiple machine learning networks process the 2D images to generate depth maps and orientation channels, enabling the system to distinguish between vessels at different depths (bifurcated vs. overlapping) while maintaining the simplicity of the original 2D imaging process.
Solution Approach 2:
The patent introduces intermediate representation layers including depth image channels, orientation channels, and probability maps as mediators between the input 2D medical images and the final vessel centerline extraction. These intermediate representations enable the system to differentiate vessel types without directly complicating the original imaging process.
2Reliability
If conventional centerline tracing methods are used, then the computation is fast, but false path tracing occurs due to inability to distinguish vessel types
Solution Approach 1:
The patent performs preliminary processing by generating depth image channels, orientation channels, and probability maps before performing centerline tracing. These pre-computed representations provide the necessary information for accurate vessel distinction, allowing the subsequent centerline tracing to be both reliable and efficient without requiring complex real-time analysis.
Solution Approach 2:
The patent replaces conventional mechanical/mathematical centerline tracing algorithms with machine learning-based approaches. Neural networks process the medical images to generate depth and orientation information, substituting traditional computational methods with AI-based systems that achieve higher accuracy while maintaining processing efficiency.
3Measurement precision
If machine learning networks are used for 3D depth reconstruction, then vessel distinction accuracy is improved, but the system complexity increases
Solution Approach 1:
The patent segments the complex machine learning processing into separate functional modules: a first network for depth image channel generation, a second network for orientation channel generation, and a third network for probability map generation. This segmentation allows each network to be optimized independently and simplifies the overall system architecture while maintaining high accuracy in vessel distinction.
Data Source
AI summary
Systems and methods are provided for three dimensional depth reconstruction of vessels in two dimensional medical images. A medical image comprising braches of one or more vessels is received. A branch overlap image channel that represents a pixelwise probability that the branches overlap is generated. A set of branch orientation image channels are generated. Each branch orientation image channel is associated with one of a plurality of orientations. Each branch orientation image channel representing a image channel represents a pixelwise probability that the branches are oriented in its associated orientation. A multi-channel depth image is generated based on the branch overlap image channel and the set of branch orientation image channels. Each channel of the multi-channel depth image comprises portions of the branches corresponding to a respective depth.