Cascade Brain Tumor Segmentation for Coarse-to-Fine 3D U-Net
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation methods for brain tumors, particularly using fully convolutional neural networks, suffer from poor feature learning abilities, leading to inadequate segmentation results.
Innovation Solution
An image segmentation method utilizing a cascaded segmentation network with multiple levels of segmentation networks based on machine learning models, including tumor localization and 3D U-Net networks, to perform stepwise image segmentation of brain tumors, improving segmentation accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If fully convolutional neural network method is used for tumor image segmentation, then the segmentation process can be automated, but the feature learning ability for the full image is poor leading to poor segmentation effect
Solution Approach 1:
The patent divides the image segmentation task into multiple levels: a first segmentation network performs coarse segmentation to identify general tumor regions, and a second segmentation network performs fine segmentation to identify specific tumor subregions. This multi-level segmentation approach enables the system to learn features at different granularities, improving overall segmentation accuracy while maintaining automation.
Solution Approach 2:
The patent introduces a hierarchical dimension to the segmentation process by creating multiple levels of segmentation networks. The first-level network operates on the entire image to capture global tumor characteristics, while the second-level network operates on localized regions to capture detailed subregion characteristics. This dimensional hierarchy resolves the contradiction by enabling both full-image context understanding and precise local feature learning.
2Measurement precision
If multi-level cascaded segmentation network is used to improve segmentation accuracy, then segmentation granularity is improved, but device complexity increases
Solution Approach 1:
The patent segments the complex segmentation task into two manageable levels: a first segmentation network for coarse tumor region identification and a second segmentation network for fine tumor subregion classification. Each network has a specific, focused function, which simplifies the overall system design compared to a single monolithic network attempting to perform all segmentation tasks simultaneously.
Solution Approach 2:
The first segmentation network performs preliminary coarse segmentation to identify general tumor regions before the second segmentation network performs fine segmentation on those specific regions. This preliminary action reduces the search space and computational burden for the second network, effectively managing complexity while achieving high segmentation granularity.
Data Source
Figure 1~2
Figure 3~4(c)
Figure 5~7a
AI summary
Disclosed in the present application are an image segmentation method and apparatus, a diagnosis system, a storage medium, and a computer device. The image segmentation method comprises: obtaining a tumor image; performing tumor localization on the obtained tumor image to obtain a candidate image for indicating the position of a whole tumor region in the tumor image; inputting the candidate image to a cascade segmentation network constructed on the basis of a machine learning model; and perform image segmentation on the whole tumor region in the candidate image by using a first-stage segmentation network in the cascade segmentation network as a starting point, and stepping to the last-stage segmentation network stage by stage to perform image segmentation on an enhanced tumor core region to obtain a segmented image.