3D Medical Image Semantic Segmentation via Adaptive Multi-Axis Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semantic segmentation technologies, such as Pspnet, are limited to two-dimensional medical images and cannot effectively perform segmentation on three-dimensional medical images, leading to faults and inadequate edge fitting.
Innovation Solution
A method involving three-dimensional semantic segmentation using three two-dimensional segmentation models and an adaptive fusion model to process three-dimensional medical images, incorporating slicing, two-dimensional segmentation, and three-dimensional fusion to suppress background noise and accurately segment edges.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Pspnet is used for semantic segmentation, then two-dimensional medical images can be segmented, but three-dimensional medical images cannot be segmented
Solution Approach 1:
The patent transforms the three-dimensional medical image into multiple two-dimensional slice images along different axes (x-axis, y-axis, z-axis), enabling the application of two-dimensional segmentation models to three-dimensional data. This dimensional decomposition resolves the contradiction by making the system adaptable to three-dimensional inputs while maintaining reliability through proven two-dimensional segmentation algorithms.
Solution Approach 2:
The patent divides the three-dimensional image segmentation task into multiple two-dimensional segmentation sub-tasks. By slicing the three-dimensional image into two-dimensional slices and processing each slice independently through separate segmentation models, the system achieves both adaptability to three-dimensional data and reliability through established two-dimensional segmentation techniques.
2Device complexity
If two-dimensional segmentation models are used directly on three-dimensional images, then processing is simpler, but edge fitting is inadequate and noise increases
Solution Approach 1:
The patent merges the results from multiple two-dimensional segmentation models through a fusion module. By combining the segmentation results from x-axis, y-axis, and z-axis models, the system achieves superior edge fitting and noise suppression while maintaining relatively simple individual model structures. The fusion process integrates multiple perspectives to produce high-precision three-dimensional segmentation results.
3Manufacturing precision
If three-dimensional fusion is performed on distribution probability maps, then noise is suppressed and edges are smoothed, but computational complexity increases
Solution Approach 1:
The patent introduces distribution probability maps as intermediary representations between raw image data and final segmentation results. The fusion module operates on these probability maps rather than raw images, enabling effective noise suppression and edge smoothing while managing computational complexity through probabilistic integration of multiple segmentation perspectives.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
This application belongs to the field of deep learning, and discloses a semantic segmentation method and apparatus for a three-dimensional image, a terminal, and a storage medium. The method includes: obtaining a three-dimensional image; performing slicing on the three-dimensional image according to three directional planes in which three-dimensional coordinate axes are located, to obtain a two-dimensional slice image of an x axis, a two-dimensional slice image of a y axis, and a two-dimensional slice image of a z axis; invoking a first segmentation model, a second segmentation model, and a third segmentation model to perform semantic segmentation on two-dimensional slice images of the x axis, the y axis, and the z axis, to obtain distribution probability maps of a target object on the three directional planes; and invoking an adaptive fusion model to perform three-dimensional fusion on the three distribution probability maps respectively corresponding to an x-axis directional plane, a y-axis directional plane, and a z-axis directional plane, to obtain a three-dimensional distribution binary image of the target object.