Knowledge Distillation for Low-Noise Medical Image Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging techniques suffer from high noise levels, leading to low signal to noise (SNR) and contrast to noise (CNR) ratios, which degrade image quality and hinder accurate diagnosis and treatment.
Innovation Solution
A knowledge distillation based training method using a cGAN machine learning model to generate a reduced noise dataset, followed by training a UNET model on this dataset, to enhance image quality and reduce computational burdens.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If current medical imaging techniques are used, then image scanning capability is improved, but signal to noise ratio and contrast to noise ratio deteriorate
Solution Approach 1:
The patent applies preliminary action by training a first machine learning model in advance using a first dataset with high noise and a second dataset with low noise. This pre-trained model then generates a target dataset with reduced noise, which is subsequently used to train a target machine learning model. This two-stage preliminary preparation enables the final model to achieve high SNR and CNR while maintaining fast processing speed.
Solution Approach 2:
The patent introduces an intermediary target dataset generated by the first machine learning model. This target dataset acts as a bridge between the noisy first dataset and the final target model. The intermediary dataset contains synthesized images with reduced noise levels, enabling the target model to learn noise reduction patterns without directly processing extremely noisy data.
2Measurement precision
If noise reduction is applied to improve image quality, then signal to noise ratio and contrast to noise ratio are improved, but computational processing time increases
Solution Approach 1:
The patent uses copying by having the first machine learning model generate synthetic target images that replicate the characteristics of low-noise reference images. Instead of performing complex real-time noise reduction on every input image, the system creates a copy of the desired low-noise output pattern through the trained first model, which then guides the target model to produce similar results efficiently.
Solution Approach 2:
The patent applies parameter changes by transforming the noise characteristics of images through the machine learning models. The first model learns to map high-noise parameters to low-noise parameters, and the target model uses this learned transformation to rapidly adjust noise levels in new images, achieving CNR improvement with minimal processing time.
3Measurement precision
If a first machine learning model is trained to generate reduced noise images, then noise values are reduced, but device complexity increases
Solution Approach 1:
The patent segments the noise reduction task into two distinct phases: first, training a generation model to create a target dataset with reduced noise characteristics; second, training a target model using this pre-prepared target dataset. This segmentation allows each model to focus on a specific aspect of noise reduction, simplifying the overall complexity compared to training a single comprehensive model.
Solution Approach 2:
The first machine learning model performs preliminary noise reduction to create the target dataset before the target model is trained. This preliminary action prepares clean training data in advance, allowing the target model to learn from high-quality examples without needing to develop complex noise filtering capabilities itself, thereby reducing its complexity.
Data Source
AI summary
At least a method for training a target machine learning model for enhancing a digital image processing is provided. The method comprises receiving a first data set including a first plurality of digital images, training a first machine learning model using the first data set and a second data set including a second plurality of digital images, generating, by the first machine learning model that is trained, a target data set including a third plurality of digital images, the third plurality of digital images having noise represented by respective noise values that are lower than the noise represented by the respective noise values of the first plurality of digital images, and training the target machine learning model using the target data set and the first data set including the first plurality of digital images for enhancing at least one characteristic of a new digital image.