Low-Dose Myocardial Scar Segmentation Using Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for scar segmentation in medical imaging using gadolinium-based contrast agents (GBCA) face challenges in achieving accurate and reliable assessment while minimizing the amount of contrast agent used, due to safety concerns and potential gadolinium retention in the body.
Innovation Solution
A computer-implemented method for training a convolutional neural network using low-dose contrast agent images and ground truth data from full-dose images to determine scar segmentation, adjusting neural network weights for accurate scar segmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If the dose of gadolinium-based contrast agent is reduced to enhance patient safety and reduce gadolinium retention, then patient safety is improved, but the accuracy and reliability of scar segmentation deteriorates
Solution Approach 1:
A deep learning model is introduced as an intermediary between low-dose contrast agent images and accurate scar segmentation results. The model learns to map low-dose images to full-dose segmentation outcomes, effectively bridging the gap between reduced contrast dosage and maintained segmentation accuracy.
Solution Approach 2:
The invention changes the parameter of contrast agent dose from standard levels to reduced levels, while using deep learning to compensate for the resulting image quality degradation. This allows operating in a previously unacceptable parameter regime (low dose) while achieving the desired outcome (accurate segmentation).
2Measurement precision
If standard dose of contrast agent is used to ensure accurate scar segmentation, then measurement precision is improved, but patient safety and reduction of gadolinium retention worsen
Solution Approach 1:
The deep learning model serves as an intermediary that processes low-dose images to produce segmentation results equivalent to those obtained from high-dose images, eliminating the need to administer harmful levels of contrast agent.
Solution Approach 2:
The model is trained by copying ground truth segmentation masks from full-dose images to guide the segmentation process on low-dose images, allowing the system to learn the mapping without requiring patients to receive high doses of contrast agent during actual use.
3Measurement precision
If deep learning model is trained with low-dose images and full-dose ground truth masks, then accurate segmentation with reduced contrast dose is achieved, but training data requirements and processing complexity increase
Solution Approach 1:
The deep learning model is trained in advance on paired datasets of low-dose images and corresponding full-dose ground truth masks. This preliminary training phase creates a pre-configured model that can directly process new low-dose images without requiring complex processing during actual deployment.
Solution Approach 2:
The training process uses feedback from comparing model predictions against ground truth full-dose segmentation masks to iteratively improve model weights, ensuring the model learns accurate mappings from low-dose images to segmentation results.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
A computer-implemented method for determining scar segmentation, in particular segmentation of a myocardial scar, comprising the following steps: Receiving a medical image (5) of an object to be segmented acquired after an application of a low-dose of contrast agent, wherein the low-dose of contrast agent comprises less contrast agent than a standard full-dose of contrast agent; determining a scar segmentation mask (44) by applying a trained artificial neural network (34), to the medical image.