Multi-task Coronary Artery Segmentation for Stenosis Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for assessing coronary artery disease (CAD) through coronary CTA imaging struggle with accurately identifying stenoses, especially when plaque buildup is diffuse or occurs at bifurcations, and AI-based systems face limitations due to reliance on diameter or area profiles, leading to reduced accuracy.
Innovation Solution
A machine learning-based model using multi-task learning with shared features for segmentation of reference walls and lumens, incorporating regularization for consistency and uncertainty estimation, enables accurate stenosis grading and vessel assessment tasks, robust to image artifacts and poor quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional interpolation or iterative fitting methods are used to estimate healthy coronary artery diameter or area, then the assessment process is simple and fast, but the accuracy is reduced when plaque buildup is diffuse, long, or occurs at bifurcations
Solution Approach 1:
The patent segments the coronary artery assessment into multiple independent tasks: lumen segmentation, reference wall segmentation, and stenosis grading. Each task has its own dedicated decoder module that processes shared encoder features independently, allowing accurate segmentation even in complex cases with diffuse plaque or bifurcations while maintaining computational efficiency through feature sharing.
Solution Approach 2:
The patent implements a multi-task learning framework where a single encoder network serves multiple segmentation and assessment tasks simultaneously. The shared encoder extracts common features that are then distributed to multiple decoders for different assessment objectives, improving overall system efficiency and consistency while maintaining high accuracy across all tasks.
2Measurement precision
If AI-based systems are trained using only diameter or area profile of the lumen, then the training data requirement is reduced, but the accuracy of stenosis assessment is reduced
Solution Approach 1:
The patent separates the assessment into distinct segmentation tasks for lumen and reference walls, each with its own decoder. This allows the system to capture detailed geometric information from both structures independently, providing more comprehensive features for stenosis assessment beyond simple diameter or area profiles while maintaining structured processing efficiency.
Solution Approach 2:
The patent combines multiple segmentation tasks (lumen segmentation, reference wall segmentation) and stenosis grading into a unified multi-task learning framework. The shared encoder merges common features from all input images, while separate decoders and grading modules process task-specific information, achieving high accuracy through integrated multi-task processing.
3Stability of the object's composition
If conventional single-task approaches are used for vessel assessment, then the model training is simpler, but the consistency between different assessment results is reduced
Solution Approach 1:
The patent merges multiple assessment tasks (lumen segmentation, reference wall segmentation, stenosis grading) into a single unified multi-task learning framework. All tasks share the same encoder and are trained simultaneously with a combined loss function, ensuring that results from different tasks are consistent and mutually supportive while capturing interdependencies between segmentation and grading.
Solution Approach 2:
The patent implements a universal encoder that serves all assessment tasks simultaneously, extracting features that are relevant to multiple objectives. This shared feature representation ensures consistency across different assessment results while the task-specific decoders and grading modules maintain the necessary specialization for each function.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and methods for automated assessment of a vessel are provided. One or more input medical images of a vessel of a patient are received. A plurality of vessel assessment tasks for assessing the vessel is performed using a machine learning based model trained using multi-task learning. The plurality of vessel assessment tasks comprises segmentation of reference walls of the vessel from the one or more input medical images and segmentation of lumen of the vessel from the one or more input medical images. Results of the plurality of vessel assessment tasks are output.