Carotid Plaque Segmentation via Two-Stage Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated segmentation methods for atherosclerotic plaque components in multi-weighted MR images are prone to inter and intra-reader variability and are highly dependent on manually provided reference values, which are error-prone due to limitations in training set image quality and resolution.
Innovation Solution
A two-stage neural-network-based method using a convolutional neural network (CNN) for inner and outer vessel wall segmentation and a Bayesian deep neural network (BNN) for pixel-level segmentation of plaque components, trained with high-resolution ex vivo MR images and histopathology to refine manually-defined in vivo MR image training sets, reducing variability and improving segmentation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated segmentation methods are used for atherosclerotic plaque components, then segmentation speed is improved, but measurement precision deteriorates due to inter and intra-reader variability
Solution Approach 1:
The patent divides the segmentation task into two distinct stages: a first-stage model that segments the vessel wall and lumen, and a second-stage model that segments plaque components within the vessel wall. This multi-stage segmentation approach improves both efficiency and precision by breaking down the complex task into manageable steps, with each stage focusing on specific anatomical structures.
Solution Approach 2:
The first-stage segmentation model acts as an intermediary that provides preliminary annotations to guide the second-stage model. This intermediate step reduces variability by providing a consistent reference framework, allowing the second stage to focus specifically on plaque component segmentation with improved accuracy.
2Ease of manufacture
If manually provided reference values are used for training, then ease of manufacture is improved, but measurement precision deteriorates due to error-prone manual annotations
Solution Approach 1:
The system uses semi-automated annotation where the first-stage model automatically segments the vessel wall and lumen, eliminating the need for manual annotation of these structures. This self-service approach maintains ease of data preparation while significantly improving annotation accuracy, as the automated model provides consistent references without human error.
3Device complexity
If traditional single-stage segmentation models are used, then device complexity is reduced, but measurement precision deteriorates for plaque component segmentation
Solution Approach 1:
The patent implements a two-stage neural network model where the first stage focuses on vessel wall and lumen segmentation, and the second stage specializes in plaque component segmentation. This segmentation of the model architecture improves plaque component accuracy by dedicating specific network capacity to each task, while the modular structure keeps complexity manageable.
Solution Approach 2:
Each stage of the model is optimized for its specific function: the first-stage model is optimized for vessel wall and lumen boundaries, while the second-stage model is optimized for plaque component characteristics. This local optimization of model quality for different anatomical structures improves overall segmentation precision without requiring excessive complexity.
Data Source
AI summary
Disclosed are systems, apparatuses, processes, and computer-readable media to provide haptic feedback in electronic devices based on contextual awareness. A method of processing image data includes receiving at least one image of a patient; providing the image of the patient to a trained model for segmenting the image; and processing the image based on a multi-stage trained model to segment the image into a plurality of segmented images that are separated based on plaque components.