Medical Image Compression via Split Neural Network Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional medical image data compression technologies fail to maintain a desired level of image definition after expansion, often resulting in insufficient compression or deterioration of image quality.
Innovation Solution
A medical information processing system utilizing two trained models, where a first processor compresses medical image data using a model with an input layer and a middle layer, and a second processor expands the data using an output layer from the same trained models, ensuring appropriate compression while maintaining image definition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional compression processing is applied to medical image data, then data quantity is reduced for storage efficiency, but the definition of the expanded image deteriorates
Solution Approach 1:
The trained model is divided into two separate models: a compression model (input layer + middle layer) and an expansion model (middle layer + output layer). This segmentation allows specialized optimization for both compression and expansion tasks, resolving the contradiction by enabling high compression ratios while maintaining image definition through coordinated training of both models.
Solution Approach 2:
The patent changes the parameters of the neural network models by dividing a pre-trained model into two models with different architectures and training objectives. The compression model is trained to minimize loss in the compressed domain, while the expansion model is trained to reconstruct high-quality images, allowing parameter optimization for both compression efficiency and image quality simultaneously.
2Use of energy by moving object
If compression ratio is increased to reduce storage resources, then resource consumption is reduced, but image definition after expansion deteriorates
Solution Approach 1:
The expansion model receives feedback from the compressed intermediate data and adjusts its reconstruction parameters to maintain image definition. The coordinated training process establishes a feedback mechanism where the compression model's output is continuously optimized by the expansion model to preserve image quality even at high compression ratios, resolving the resource consumption vs. image definition contradiction.
Solution Approach 2:
The system performs preliminary compression using the compression model to reduce data quantity and resource consumption, then applies preliminary expansion using the expansion model to restore image definition before final output. This preliminary action sequence ensures both resource efficiency and image quality are achieved.
Data Source
AI summary
A first processor of embodiments outputs intermediate data with a quantity less than that of third medical image data by inputting the third medical image data to a compression model including an input layer and a middle layer from two trained models obtained by dividing, on the basis of the middle layer, a trained model which has been trained such that second medical image data is output from an output layer by inputting first medical image data to the input layer. A second processor outputs fourth medical image data with a quantity greater than that of the intermediate data by inputting the intermediate data acquired from the first processor via a network to an expansion model including the output layer from the two trained models.