3D Convolution Neural Network Subnetworks for Medical Image Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep convolutional neural networks (CNNs) face significant computational resource requirements, making them inefficient for processing large 3D medical images, particularly in scenarios with limited memory and computational capacity, which restricts their application in mobile and embedded computing.
Innovation Solution
The implementation of a 3D deep convolutional neural network architecture (DCNNA) with subnetworks that perform dimensionality reduction and multi-scale convolutions, combined with 3D batch normalization, to enhance computational efficiency and learning rates, allowing for more effective processing of 3D data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional CNNs process 3D data with multiple convolutional layers, then recognition performance is improved, but computational resource requirements increase significantly
Solution Approach 1:
The patent divides the convolutional processing into separate paths: a first path performs standard 3D convolutions for feature extraction, while a second path performs 2D convolutions on projected 2D slices. This segmentation allows the system to achieve recognition performance through multiple processing paths while reducing overall computational resource requirements by avoiding redundant 3D convolution operations.
Solution Approach 2:
The patent transforms 3D volumetric data into 2D projections and processes them through 2D convolutional layers. This dimensionality change enables the system to maintain recognition capability through multi-scale feature extraction while significantly reducing computational complexity compared to traditional 3D-only CNN approaches.
2Measurement precision
If 3D deep convolutional neural networks process large 3D medical images, then image classification accuracy is improved, but memory and computational capacity requirements increase
Solution Approach 1:
The patent segments the processing of 3D medical images into 2D slice processing and 3D volumetric processing paths. The 2D path handles local detailed features with lower computational cost, while the 3D path handles global contextual features. This segmentation enables high classification accuracy on large 3D medical images while reducing overall memory and computational capacity requirements.
Solution Approach 2:
The patent applies different processing qualities to different regions: 2D convolutions are applied to 2D projections of 3D slices for local feature extraction, while 3D convolutions are applied for global feature integration. This local quality differentiation allows the system to achieve high accuracy on complex 3D medical images while optimizing computational resource usage.
3Speed
If subnetworks perform multi-scale convolutions in parallel, then processing speed is improved, but network architecture complexity increases
Solution Approach 1:
The patent segments the convolutional network into distinct first and second subnetworks with different convolutional paths. The first subnetwork handles 3D convolutions while the second handles 2D convolutions on projected data. This segmentation enables parallel multi-scale processing that increases speed while keeping architecture complexity manageable through clear functional separation.
Solution Approach 2:
The patent introduces a 2D projection dimension alongside the 3D volumetric dimension. By processing 2D projections in parallel with 3D data through separate convolutional paths, the system achieves faster multi-scale processing while the added 2D processing dimension provides complementary feature extraction without proportionally increasing overall architecture complexity.
Data Source
AI summary
A method of classifying three-dimensional (3D) data includes receiving three-dimensional (3D) data and processing the 3D data using a neural network that includes a plurality of subnetworks arranged in a sequence and the data is processed through each of the subnetworks. Each of the subnetworks is configured to receive an output generated by a preceding subnetwork in the sequence, process the output through a plurality of parallel 3D convolution layer paths of varying convolution volume, process the output through a parallel pooling path, and concatenate output of the 3D convolution layer paths and the pooling path to generate an output representation from each of the subnetworks. Following processing the data through the subnetworks, the method includes processing the output of a last one of the subnetworks in the sequence through a vertical pooling layer to generate an output and classifying the received 3D data based upon the generated output.