Adaptive Pattern Transformation for Robust Deep Neural Reproduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural network models exhibit significant performance variations when applied to medical images with different qualitative patterns due to differences in manufacturer preferences, professional imaging styles, racial differences, and capture environments, leading to inefficient and costly customization for each institution and uncertainty in data quality.
Innovation Solution
A method that retrieves candidate data with the highest similarity to target data from a learning data representative group, performs adaptive pattern transformation to align the target data with the candidate data, and transfers the transformed data to a deep neural network model to maintain consistent performance across varying qualitative patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a deep neural network model is trained on training data with one qualitative pattern, then it achieves good performance on that pattern, but performance degrades significantly when applied to target data with different qualitative patterns
Solution Approach 1:
The patent transforms the target data's qualitative parameters to match the training data's qualitative pattern through adaptive pattern transformation. This involves changing parameters such as image style, texture, and appearance characteristics while preserving the underlying semantic content, enabling the model to process diverse data patterns reliably
Solution Approach 2:
The patent introduces an intermediate transformation process that acts as a mediator between the target data and the trained model. This transformation module converts target data with unknown or different qualitative patterns into a format consistent with the training data, bridging the gap between diverse inputs and the fixed-model requirements
2Reliability
If deep neural network models are customized for each institution and country with different qualitative patterns, then performance on specific patterns improves, but the cost and complexity increase significantly
Solution Approach 1:
The patent creates a universal preprocessing system that can handle multiple qualitative patterns through a single adaptive transformation module. This one-size-fits-all approach replaces the need for multiple customized models, reducing complexity while maintaining high performance across different institutions and countries
Solution Approach 2:
The transformation module automatically adapts target data to match training data patterns without requiring manual intervention or customization. The system self-adjusts by retrieving reference data with similar qualitative patterns and performing automatic transformation, eliminating the need for expensive manual quality matching
3Reliability
If manual quality matching is performed for each target data to ensure consistent pattern, then reproduction performance improves, but the time and computational resources required increase
Solution Approach 1:
The system performs automatic pattern transformation by retrieving reference data and applying adaptive transformations without manual intervention. The transformation module autonomously matches target data with appropriate reference patterns and executes the transformation, eliminating time-consuming manual quality matching while maintaining output consistency
Solution Approach 2:
The patent pre-processes target data by performing adaptive pattern transformation before feeding it to the trained model. This preliminary transformation ensures that the data is in the correct format and pattern, preventing performance degradation and eliminating the need for post-processing or manual adjustment
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to a method for improving reproduction performance of a deep neural network model trained using a group of learning data so that the deep neural network model can exhibit excellent reproduction performance even for target data having a quality pattern different from that of the group, and a device using same. According to the method of the present disclosure, a computing device acquires the target data, retrieves at least one piece of candidate data having a highest similarity to the target data from a learning data representative group including reference data selected from the learning data, performs adaptive pattern transformation on the target data to enable adaptation to the candidate data, and supports transfer of transformed data, which is a result of the adaptive pattern transformation, to the deep neural network model so as to acquire an output value from the deep neural network model.