Learning-Based Active Surface Model for 3D Medical Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image segmentation methods, particularly for 3D tasks like segmenting small organs from CT scans, face challenges with high variability in shape, size, and location, leading to over-segmentation and inaccurate boundaries due to mismatched input and output data in traditional deep learning approaches.
Innovation Solution
A learning-based active surface model comprising a surface initialization network and a surface evolution network that uses point cloud data to deform the initial surface for accurate 3D segmentation, improving both accuracy and efficiency by leveraging deep neural networks for fully automated segmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If 2D networks are used to process 3D volumetric data by slicing, then computational cost is reduced, but 3D geometric information is lost leading to inaccurate boundaries
Solution Approach 1:
The patent transitions from processing 2D slices to processing 3D volumetric data directly by representing surfaces as point clouds in 3D space. This dimensional transformation allows the network to preserve and utilize 3D geometric information while maintaining computational efficiency through point-based representation rather than full volumetric processing.
Solution Approach 2:
The patent extracts the essential geometric information from 3D volumetric data by representing surfaces as point clouds. This extraction approach captures the critical boundary information needed for segmentation while discarding redundant internal volume data, thereby reducing computational cost while maintaining boundary accuracy.
2Measurement precision
If 3D networks are used to process volumetric data directly, then 3D geometric information is preserved, but GPU memory consumption increases significantly
Solution Approach 1:
The patent extracts only the essential surface information from 3D volumetric data by representing boundaries as point clouds. This selective extraction maintains 3D geometric accuracy while dramatically reducing the data volume that needs to be processed and stored in GPU memory, solving the memory bottleneck of full 3D network processing.
Solution Approach 2:
The patent segments the 3D volumetric data by focusing computation only on the surface points that define organ boundaries, rather than processing the entire volume. This segmentation approach preserves 3D geometric information at the critical boundary regions while avoiding the memory overhead of processing all volumetric pixels.
3Quantity of substance
If traditional active surface methods are used, then memory consumption is low, but convergence speed is slow and initialization is complex
Solution Approach 1:
The patent replaces the traditional mechanical active surface evolution process with a data-driven neural network approach. Instead of iteratively evolving surfaces through numerical optimization, the network directly predicts surface points from volumetric data, eliminating the slow convergence process while maintaining low memory consumption through point cloud representation.
Solution Approach 2:
The patent enables the segmentation system to automatically initialize and evolve surfaces through learned patterns from training data. The network self-adjusts to produce accurate surface predictions without requiring manual initialization points or complex topological constraints, thereby improving productivity while keeping the system memory-efficient.
4Extent of automation
If learning-based active surfaces use 2D U-nets to estimate surfaces from image intensities, then automation is improved, but prediction accuracy deteriorates due to input-output data mismatch
Solution Approach 1:
The patent resolves the input-output data mismatch by changing the dimensionality of the output from 2D intensity-based predictions to 3D point cloud representations. This dimensional transformation allows the network to work directly with spatial coordinates of surface points, improving prediction accuracy while maintaining full automation through the learned mapping from volumetric data to surface points.
Data Source
AI summary
A learning-based active surface model for medical image segmentation uses a method including: (a) data generation: obtaining medical images and associated ground truths, and splitting the sample images into a training set and a testing set; (b) raw segmentation: constructing a surface initialization network, parameters of the network trained by images and labels in the training set; (c) surface initialization: segmenting the images by the surface initialization network, and generating the point cloud data as the initial surface from the segmentation; (d) fine segmentation: constructing the surface evolution network, the parameters of the network trained by the initial surface obtained in step (c); (e) surface evolution: deforming the initial surface points along the offsets to obtain the predicted surface, the offsets presenting the prediction of the surface evolution network; (f) surface reconstruction: reconstructing the 3D volumes from the set of predicted surface points set to obtain the final segmentation results.