CT Perfusion Segmentation Using 4D Data and Parameter Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for segmenting the core and penumbra in ischemic stroke lesions using computed tomography perfusion datasets lack reliable ground truth data and fail to adequately account for the correlation between these regions, leading to inconsistent and unreliable segmentation.
Innovation Solution
A computer-implemented method that utilizes both perfusion parameter maps and the raw 4D CTP dataset to train a deep learning model, employing a multi-stage neural network architecture with separate subfunctions for penumbra and core prediction, guided by follow-up examination data and success indicators to enhance segmentation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If threshold segmentation is applied based on perfusion parameter maps, then segmentation can be performed, but the reliability and accuracy of segmentation is insufficient due to lack of standardized thresholds and ground truth data
Solution Approach 1:
The patent introduces an artificial intelligence model as an intermediary between the perfusion parameter maps and the segmentation result. The AI model learns the complex mapping relationship from training data and serves as a mediator that transforms the perfusion parameters into accurate segmentation maps, resolving the issue of unreliable threshold-based segmentation.
Solution Approach 2:
The patent creates a virtual copy of the ground truth segmentation through AI prediction. Instead of relying on manual thresholding, the system generates segmentation maps by copying the essential features from training data through the AI model, enabling reliable segmentation without requiring standardized manual thresholds.
2Measurement precision
If existing segmentation methods are used, then processing time is relatively short, but the accuracy and consistency of core and penumbra distinction is insufficient
Solution Approach 1:
The patent segments the segmentation task into two distinct parts by using separate AI models or separate prediction stages for core and penumbra. This allows each region to be optimized independently with region-specific thresholds and validation criteria, improving overall segmentation accuracy while maintaining manageable complexity.
Solution Approach 2:
The patent changes the approach from fixed threshold parameters to learned parameters through AI training. The system transforms the segmentation problem from using static, manually-determined thresholds to using dynamic, data-driven parameters that adapt to different patient cases, significantly improving accuracy.
3Reliability
If manual threshold selection is used, then method simplicity is maintained, but segmentation consistency and reliability deteriorate due to personal preferences and lack of standardization
Solution Approach 1:
The patent implements self-service segmentation where the AI model automatically selects optimal thresholds and parameters based on the input perfusion data. The system serves itself by learning from training data and making autonomous segmentation decisions without requiring manual threshold selection, ensuring consistency while maintaining ease of use through automated operation.
Solution Approach 2:
The patent incorporates feedback mechanisms where the AI model is trained on ground truth data and continuously improves its segmentation performance. The system uses validation against known outcomes to adjust its parameters, creating a feedback loop that enhances reliability and consistency while eliminating the need for manual intervention.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
Computer-implemented segmentation method for segmenting a core (52) and a penumbra (53) in a four-dimensional computed tomography perfusion dataset (38) of ischemic tissue in an image region of a patient, wherein at least one parameter map (39) for at least one perfusion parameter is determined from the computed tomography perfusion dataset (38), whereafter a trained function (60) using the at least one parameter map (39) as input data determines output data comprising segmentation information (42) of the penumbra (53) and the core (52), wherein the trained function (60) uses the computed tomography perfusion dataset (38) as additional input data.