Gated Spatiotemporal Unit for Irregular Shape Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation techniques, such as region growing and CNNs, struggle to accurately identify and segment irregularly-shaped objects in medical images due to their reliance on pixel adjacency and regularity, leading to incorrect boundary identification and failure in recognizing shapes with varying geometry and intensity.
Innovation Solution
A gated spatiotemporal unit is introduced, combining principles of RNNs and CNNs to propagate information over both time and space, allowing nodes to update their internal states based on previous states and neighboring nodes, and utilizing an image pyramid for efficient data propagation across different resolutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If region growing is used to segment objects in medical images, then simple and fast segmentation can be achieved, but incorrect spreading occurs across bright tissue boundaries leading to poor segmentation accuracy
Solution Approach 1:
The patent introduces a learned boundary indicator function as an intermediary between the region growing process and the image data. This function acts as a gatekeeper that controls whether pixel spreading should occur across potential boundaries, preventing incorrect spreading while maintaining the efficiency of region growing. The boundary indicator is learned through training and provides adaptive control over the segmentation process.
Solution Approach 2:
The patent makes the region growing process dynamic by allowing the boundary indicator to adaptively control spreading based on learned characteristics. Instead of fixed threshold-based spreading, the system dynamically adjusts which boundaries permit spreading based on the learned boundary indicator function, enabling the system to distinguish between valid and invalid boundaries in different contexts.
2Measurement precision
If CNNs are used to identify objects with regularity, then accurate classification of regular shapes can be achieved, but irregularly-shaped objects like tumors and lesions cannot be accurately segmented
Solution Approach 1:
The patent segments the image into multiple regions through controlled pixel spreading and uses boundary indicators to classify different regions. This region-based approach allows the system to handle irregular shapes by treating each segmented region as a separate object, rather than relying on fixed geometric assumptions about object regularity.
Solution Approach 2:
The patent changes the fundamental parameters of the segmentation approach by moving from fixed geometric assumptions to learned boundary characteristics. The boundary indicator function learns adaptive parameters from training data that capture the essence of different object types, enabling accurate segmentation of irregular shapes without relying on regularity assumptions.
3Measurement precision
If sophisticated methods like level sets and graph cuts are used instead of region growing, then segmentation accuracy can be improved, but computational complexity and processing time increase significantly
Solution Approach 1:
The patent introduces a learned boundary indicator function as a simple intermediary that guides the region growing process. This intermediary provides the accuracy benefits of sophisticated methods by learning boundary characteristics, while keeping the overall algorithm simple and efficient by building upon the basic region growing framework rather than replacing it with complex optimization methods.
Solution Approach 2:
The patent copies the simple and efficient structure of region growing while enhancing it with a learned boundary indicator. Rather than adopting the computational complexity of level sets or graph cuts, the system copies the efficient region growing mechanism and augments it with a learned gating function that provides the necessary accuracy for complex medical image segmentation.
Data Source
AI summary
A system for determining a region of interest in an image. The system includes a memory and an electronic processor. The electronic processor is configured to create an image pyramid for the image. The image pyramid includes a plurality of layers. For each layer of the image pyramid, the electronic processor is configured to initialize internal states of nodes of a spatial lattice and iteratively update the internal states of the nodes using spatially gated propagation. Each node in the spatial lattice represents a block of pixels in the image and is connected to a node representing a neighboring block of pixels in the image. The electronic processor is also configured to identify the region of interest based on the internal states of the nodes at a convergence of the spatial lattice having nodes representing values included in a first layer of the image pyramid.