Tubular Structure Segmentation Using 2D Views for Faster 3D Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for segmenting tubular structures in medical images, such as bronchus trees and pulmonary arteries and veins, are time-consuming, prone to errors, and inefficient due to the complexity of the structures and the use of non-specific neural network architectures, leading to a waste of resources and lack of accuracy.
Innovation Solution
A hierarchical, cascaded computational method using 2D and 2.5D neural networks, specifically trained for body part and tubular structure segmentation, processes 2D medical images from distinct sectional views to create a labeled 3D image, optimizing resource use and reducing processing time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual or interactive segmentation tools are used to extract tubular structures, then segmentation can be performed, but the process is time-consuming and prone to error
Solution Approach 1:
The patent replaces manual/interactive mechanical segmentation processes with an automated deep learning system. The neural network automatically processes medical images to segment tubular structures, eliminating the need for manual tracing while maintaining or improving accuracy. This substitution directly addresses the contradiction by providing automated processing that is both fast and reliable.
Solution Approach 2:
The segmentation system performs self-service by automatically processing images without requiring continuous human intervention. The deep learning model independently completes the segmentation task, making the process both time-efficient and reliable through automated decision-making algorithms.
2Productivity
If standard non-specific neural network architectures are used for segmentation, then automation is achieved, but there is a waste of resources and lack of efficiency and accuracy
Solution Approach 1:
The patent applies local quality by designing a neural network architecture specifically optimized for tubular structure segmentation rather than using generic models. The architecture includes specialized components like tubular loss functions and multi-scale analysis modules that are locally adapted to the characteristics of tubular structures, improving efficiency and reducing wasted computational resources on irrelevant features.
Solution Approach 2:
The patent changes key parameters of the neural network architecture to optimize for tubular structures, including modified loss functions, adjusted convolutional kernel sizes, and optimized learning rates. These parameter changes enable the network to converge faster and use computational resources more efficiently while maintaining high accuracy.
3Measurement precision
If 3D networks are used for bronchus tree segmentation, then representation power is improved, but computation tractability becomes more difficult
Solution Approach 1:
The patent segments the 3D segmentation task into multiple 2D processing steps. Instead of processing the entire 3D volume at once with a computationally intensive 3D network, the method processes individual 2D slices through the body part, performing tubular structure segmentation on each slice independently. This segmentation approach maintains accuracy while dramatically reducing computational complexity.
Solution Approach 2:
The patent transitions from 3D processing to 2D processing by analyzing tubular structures in two-dimensional cross-sectional views. This dimensionality change simplifies the computational burden while preserving the ability to accurately identify and segment tubular structures through the use of specialized 2D neural network architectures and loss functions.
Data Source
AI summary
This invention concerns a computer implemented method for segmenting and identifying at least one tubular structure, having a 3D tree layout and located in at least one body part of a subject, in medical images showing a volume region of interest of said subject containing said body part(s), and for providing a labelled 3D image of said structure(s),said method mainly comprising the steps of:providing a set of 2D medical images corresponding to respective mutually distinct sectional views across said region of interest containing said body part(s), the planes of said medical images being all perpendicular to a given direction or all mutually intersecting at a given straight line,segmenting the visible section(s) of the concerned body part(s) present in each one of said 2D medical images, which comprises in particular the complete linear outline or external boundary of said body part(s) visible in the considered 2D image, and creating a corresponding 2D body part masking image,pre-processing each 2D medical image, by applying the corresponding body part masking image to it and so producing processed images containing only the image data of the original 2D image which are related to said body part(s),segmenting the tubular structure(s) in said resulting pre-processed images, possibly by segmenting tubular structures of different kinds in differentiated segmentation processes,performing the previous steps with at least one other set of 2D medical images corresponding to other respective distinct sectional views, along other mutually parallel or intersecting planes, of said same volume region of interest containing said same body part(s),merging the results of the tubular structure segmentations of the different sets of pre-processed images, in order to provide a labelled 3D image of said tubular structure(s) of one or different kind(s).