3D Image Slice Labeling via Neural Network Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Annotating labels for 3D images is resource-intensive, making it challenging to construct learning models for 3D image recognition, as existing methods require extensive annotation across all slices, which is inefficient and costly.
Innovation Solution
A method where a neural network is trained on annotated slices of a 3D image, and similarity-based expansion is used to identify and label unannotated slices, allowing the network to expand the annotated dataset, reducing the need for extensive annotation and resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If labels are annotated on all slices of a 3D image, then the learning model can be trained with complete supervision, but the annotation time and resource consumption increase significantly
Solution Approach 1:
The patent applies partial action by annotating labels on only some slices (first slice) rather than all slices of the 3D image. The neural network learns from the annotated slice and then identifies expandable slices (second slices) with similar characteristics, assigning them the same labels. This partial annotation approach significantly reduces annotation time while maintaining learning model effectiveness through similarity-based expansion.
2Quantity of substance
If labels are annotated on multiple slices, then the dataset size increases improving model training, but the annotation cost and resource requirements increase
Solution Approach 1:
The patent uses copying by identifying slices that are similar to the annotated first slice and copying the label information to these second slices. The neural network calculates similarity between slices and assigns labels to expandable slices based on this similarity, effectively creating additional training data without the need for manual annotation of each slice. This dramatically reduces annotation cost while increasing the effective dataset size.
3Productivity
If similarity-based expansion is used to identify unannotated slices, then annotation work is reduced, but the complexity of determining slice similarity increases
Solution Approach 1:
The patent introduces an intermediary mechanism - a neural network - that automatically calculates slice similarity and identifies expandable slices. The neural network serves as a mediator between the annotated first slice and unannotated second slices, computing feature maps and task outputs to determine similarity. This automated intermediary approach reduces manual annotation work while managing the complexity through learned feature representations rather than manual similarity assessment.
Data Source
AI summary
A 3D image sliced into a plurality of slices including the first slice on which a label is annotated and a plurality of second slices on which the label is not annotated is provided as a training sample. A computing device trains a neural network based on the first slice, determines an expandable second slice which is expandable from the first slice from among the plurality of second slices based on the trained neural network; and trains the neural network based on expanded slices including the expandable second slice.