Endoscope Learning Models with Condition-Specific Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Normalization techniques like batch normalization are not effectively applicable in hierarchical networks when data is acquired under different conditions, leading to inefficient learning.
Innovation Solution
A hierarchical network configuration with independent input layers for data groups under different conditions, accompanied by dedicated normalization layers for each group, allowing for appropriate normalization and learning across varying data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a hierarchical network with a common intermediate layer is used to process data groups acquired under different conditions, then the network structure is simplified and parameter sharing is achieved, but normalization cannot be appropriately performed on feature quantities from different data groups
Solution Approach 1:
The patent divides the normalization function into separate normalization layers for each data group (first normalization layer for first data group, second normalization layer for second data group). This segmentation allows each normalization layer to independently process feature quantities from its corresponding data group under appropriate normalization conditions, resolving the contradiction between simplified structure and normalization accuracy.
2Manufacturing precision
If separate normalization layers are provided for each data group acquired under different conditions, then appropriate normalization can be performed on feature quantities, but the network structure becomes more complex
Solution Approach 1:
The patent applies different normalization conditions locally to each data group by providing separate normalization layers (first normalization layer with first normalization condition, second normalization layer with second normalization condition). Each normalization layer is optimized for its specific data group's characteristics, achieving high normalization accuracy while maintaining a regular and systematic network structure.
3Device complexity
If normalization is performed on feature quantities from different data groups under the same conditions, then the network structure remains simple, but learning efficiency decreases due to inappropriate normalization
Solution Approach 1:
The patent segments the normalization process into separate normalization layers for each data group, allowing each layer to use normalization conditions optimized for its specific data group. This segmentation enables appropriate normalization that improves learning efficiency while maintaining a clear and organized network architecture.
Solution Approach 2:
The patent changes the normalization parameters (normalization conditions) to match the specific characteristics of each data group. By adjusting normalization parameters according to the acquisition conditions of each data group, the network achieves high learning efficiency without excessive structural complexity.
Data Source
AI summary
There are provided a learning apparatus, a learning method, a program, a trained model, and an endoscope system that can perform efficient learning even when the learning is performed using pieces of data acquired under conditions different from each other. A learning apparatus includes a hierarchical network including a first input layer, a second input layer, a first intermediate layer, a first normalization layer, a second normalization layer, a second intermediate layer, and an output layer. A learning control unit of the learning apparatus causes first learning and second learning to be performed. In the first learning, a to-be-trained model is trained based on a first error between a first recognition result and a correct answer for first data. In the second learning, the to-be-trained model is trained based on a second error between a second recognition result and a correct answer for second data.