Neural Network 3D Image Classification via 2D Slice Grouping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The scarcity of three-dimensional images poses a challenge for training machine learning models to classify such images, especially in fields where three-dimensional data is not traditionally captured, as it is difficult to obtain sufficient quantities of three-dimensional data for model training.
Innovation Solution
A neural network is trained to classify three-dimensional images by using two-dimensional slices of these images, grouped in a manner similar to how two-dimensional color images are grouped, allowing the model to accept two-dimensional spatial-channel images as input, and employing a sliding window approach to consider contextual relationships between images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models are trained to classify three-dimensional images, then classification capability for 3D data is improved, but the scarcity of three-dimensional training data worsens the training process
Solution Approach 1:
The patent segments three-dimensional images into multiple two-dimensional slices, which can be processed individually or in groups. This segmentation allows the use of abundant 2D image data to train models that can ultimately classify 3D images, effectively resolving the data scarcity problem by breaking down the 3D classification task into 2D slice processing tasks.
Solution Approach 2:
The patent transforms the problem from three-dimensional to two-dimensional by representing 3D images as sequences of 2D slices. This dimensionality reduction allows leveraging the large availability of 2D training data while maintaining the ability to classify 3D images, effectively converting a data-scarce 3D problem into a data-abundant 2D problem.
2Measurement precision
If neural networks are designed specifically for three-dimensional images, then classification accuracy for 3D data is improved, but the complexity of the model increases
Solution Approach 1:
The patent makes existing 2D image classification neural networks universal by enabling them to process 3D images through slice-based input. Instead of creating specialized 3D networks, the approach allows standard 2D networks to handle both 2D and 3D classification tasks, reducing model complexity while maintaining versatility.
Solution Approach 2:
The patent uses copies of 2D image processing architectures to handle 3D data. By applying the same proven 2D network structures to processed 3D slice sequences, the approach avoids the complexity of designing new 3D-specific architectures while achieving effective 3D image classification.
3Quantity of substance
If two-dimensional slices are used to represent three-dimensional images, then the availability of training data is improved, but the loss of spatial information worsens
Solution Approach 1:
The patent applies preliminary spatial transformations and augmentations to 2D slices before feeding them to the neural network. By pre-processing slices with spatially-aware operations that preserve contextual relationships, the approach mitigates information loss while maximizing the use of available 2D training data for 3D classification tasks.
Solution Approach 2:
The patent processes 3D images as continuous sequences of 2D slices rather than independent images. This continuous processing approach maintains spatial relationships between adjacent slices through sequential processing and contextual encoding, preserving spatial information that would be lost if slices were treated as isolated 2D images.
Data Source
AI summary
In an example embodiment, a neural network is trained to classify three-dimensional spatial-channel images in a manner that allows the training data to include two-dimensional images. Specifically, rather than redesign the neural network completely to accept three-dimensional images as input, two-dimensional slices of three-dimensional spatial-channel images are input in groupings that match the groupings that a two-dimensional image would be grouped as in the neural network. For example, if the neural network is designed to accept RGB images, it therefore is designed to accept images in groupings of three (a red component image, a green component image, and a blue component image). In such a case, the two-dimensional slices of the three-dimensional spatial-channel images will also be grouped in grouping of three so the neural network can accept them. Thus, a neural network originally designed to classify two-dimensional color images can be modified to classify three-dimensional spatial-channel images.