3D Image Processing Network Architecture Using Slice Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network architectures face limitations in processing 3D image data due to the 'curse of dimensionality,' which restricts the receptive field size and the amount of contextual information available for analysis.
Innovation Solution
A computer-implemented method that processes 3D image data by generating modified 3D image data through separate processing of 3D sub-volumes and 2D image slices using different neural networks, and then combining these modified data sets with the original 3D image data for further processing by an outputting neural network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a neural network processes 3D image data with a limited receptive field, then the processing speed and memory usage are manageable, but the amount of contextual information available for analysis is restricted
Solution Approach 1:
The patent divides the 3D image data into multiple 2D slices along different dimensions. Each slice is processed independently by neural networks with manageable receptive fields, yet the concatenation of multiple slices reconstructs the full 3D contextual information. This segmentation allows processing of large volumetric data without requiring excessively large receptive fields in single operations.
Solution Approach 2:
The patent transforms the 3D processing problem into multiple 2D processing tasks by slicing the volume along different dimensions. This dimensionality reduction allows standard 2D convolutional networks to process each slice efficiently while the combination of slices from multiple dimensions recovers the full 3D context, effectively bypassing the receptive field limitations of 3D networks.
2Measurement precision
If the receptive field size is increased to capture more contextual information, then the analysis performance improves, but the memory and processing capability requirements increase
Solution Approach 1:
By segmenting the 3D volume into 2D slices, the patent reduces the computational burden on each processing unit. Each neural network processes only a 2D slice with its associated contextual information, requiring significantly less memory and processing power than a full 3D receptive field, while maintaining analysis performance through multi-dimensional slice concatenation.
Solution Approach 2:
The patent exploits the dimensionality transformation from 3D to 2D to reduce processing requirements. 2D convolutional networks have lower computational complexity and memory demands compared to 3D networks, enabling efficient processing of medical images while preserving diagnostic quality through the reconstruction of 3D context from multiple 2D slices.
3Loss of information
If multiple 2D slices from different dimensions are processed and concatenated, then the contextual information is enhanced, but the processing time and computational steps increase
Solution Approach 1:
The patent segments the processing task into parallel 2D slice operations that can be independently computed and then concatenated. This segmentation enables parallel processing of different slices, reducing overall processing time compared to sequential 3D processing, while the concatenation of multiple dimensional slices provides enhanced contextual information.
Data Source
AI summary
A mechanism for processing input 3D image data. In a first phase, the input 3D image data is separately processed using one or more neural networks to produce one or more modified 3D image data. In a second phase, the input 3D image data and the modified 3D image data are processed using neural networks to produce an output. The neural networks that produce the modified 3D image data are configured to process slices or sub-volumes of the input 3D image data to produce modified 3D image data.