Cross-Image CNN Segmentation for PWI-DWI Stroke Penumbra Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Accurately determining the mismatch between lesions and infarcts in medical images captured using Perfusion Weighted Imaging (PWI) and Diffusion Weighted Imaging (DWI) for ischaemic stroke patients is challenging, as existing methods lack precision in identifying the penumbra region, which is crucial for reperfusion therapies.
Innovation Solution
A computer-implemented method using convolutional neural networks (CNNs) for automated segmentation, combined with region growing procedures, to generate accurate segmentation maps from PWI and DWI images, and determine the mismatch by focusing attention on regions of interest, thereby improving the accuracy of penumbra identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated segmentation using neural networks is performed on both PWI and DWI images independently, then processing speed is improved, but segmentation accuracy deteriorates due to lack of cross-image information utilization
Solution Approach 1:
The patent merges the segmentation processes by using the segmentation map from the first image type (e.g., PWI) to guide and constrain the segmentation of the second image type (e.g., DWI). This combining approach allows both images to be processed automatically while utilizing cross-image information to improve accuracy, resolving the contradiction between processing speed and segmentation accuracy.
Solution Approach 2:
The segmentation map from the first image serves as an intermediary that mediates between the two independent segmentation processes. It provides prior information and constraints that guide the second segmentation, enabling accurate penumbra region identification while maintaining automated processing efficiency.
2Measurement precision
If region growing procedures are used to generate estimated segmentation maps, then segmentation accuracy is improved, but processing time increases
Solution Approach 1:
The region growing procedure is performed in advance to generate an estimated segmentation map before the final neural network segmentation. This preliminary action provides a accurate starting point that guides subsequent processing, improving overall segmentation accuracy while the automated nature maintains reasonable processing time.
3Device complexity
If mismatch determination is performed by direct subtraction of segmentation maps, then processing simplicity is maintained, but measurement precision of penumbra region deteriorates
Solution Approach 1:
The estimated segmentation map from region growing serves as an intermediary that improves the precision of mismatch determination. By using this intermediate result to guide the neural network segmentation and mismatch calculation, the method achieves more accurate penumbra region detection while maintaining automated processing.
Data Source
Figure 1~2
Figure 3~5
Figure 6~7a
AI summary
According to an aspect, there is provided a computer-implemented segmentation method (100, 210), the method comprising: performing a first automated segmentation operation (400) on one or more first images of a subject area to automatically determine a first segmentation map of the subject area, wherein the one or more first images are generated using a first technique; performing, at least partially based on the first segmentation map, a second automated segmentation operation (600) on one or more second images of the subject area to automatically determine a second segmentation map of the subject area, wherein the one or more second images of the subject area are generated using a second technique different from the first technique, the first and second imaging techniques to capture different properties of the subject area; automatically determining a mismatch between segmented portions of the first and second segmentation maps.