Contrastive Learning Model for X-ray Image Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning algorithms face challenges in controlling image quality improvement and tend to include unwanted features, particularly in reducing noise and blurriness in medical images, making it difficult to achieve clear and accurate image restoration.
Innovation Solution
A medical image processing method using contrastive learning, where a trained model is trained with positive and negative label data to differentiate and weight specific features, improving image quality by excluding unwanted image features through a combination of positive and negative loss terms, and applying this to X-ray images to enhance image clarity and reduce noise and blurriness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning algorithms are used to improve image quality by reducing noise and blurriness, then image quality is improved, but the algorithms tend to include unwanted features and become difficult to control
Solution Approach 1:
The loss function is segmented into multiple components: a positive loss term (L1 or L2 loss) that pulls predictions toward desired image quality, and a negative loss term (contrastive loss) that pushes predictions away from unwanted features. This segmentation allows independent control over different aspects of image restoration, resolving the contradiction between improving image quality and maintaining controllability.
Solution Approach 2:
The invention changes the parameter space by introducing contrastive embeddings and using contrastive loss as an additional optimization target. By modifying the loss function parameters to include both positive and negative terms, the system gains finer control over the image restoration process, enabling it to achieve high image quality while excluding unwanted features through parameter adjustment.
2Measurement precision
If deep learning algorithms aggressively reduce noise and blurriness, then image clarity is improved, but the algorithms may over-smooth and lose important image details
Solution Approach 1:
The negative loss term acts as a counterweight to the positive loss term. While the positive loss term aggressively reduces noise and blurriness to improve image clarity, the negative loss term counteracts this by penalizing predictions that resemble unwanted blurry images, thereby preserving important image details and preventing over-smoothing.
Solution Approach 2:
The contrastive loss provides feedback by comparing predictions against both desired (positive) and undesired (negative) image examples. This dual feedback mechanism guides the optimization process to improve clarity while maintaining important features, as the model receives continuous feedback on both what it should achieve and what it should avoid.
3Measurement precision
If contrastive learning is used with positive and negative label data to control feature weighting, then unwanted image features are excluded, but the training complexity increases
Solution Approach 1:
The invention applies preliminary action by pre-processing images to generate contrastive embeddings and preparing positive and negative label data before training. By preparing the contrastive pairs and embeddings in advance, the actual training process becomes more straightforward, reducing training complexity while maintaining effective feature discrimination capability.
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
A medical image processing method according to an embodiment includes inputting first medical image data (402) to a trained model (404, 504, 504a, 804); and outputting, from the trained model, a medical image having an image quality higher than an image quality of the first medical image data, wherein the trained model (404, 504, 504a, 804) was trained using contrastive learning using second medical image data (508) as input data, third medical image data (304, 308, 510) and fourth medical image data (302, 306, 506) as label data, the third medical image data (304, 308, 510) being negative label data having worse image quality than the fourth medical image data (302, 306, 506), and the fourth medical image data (302, 306, 506) being positive label data having better image quality than the second medical image data (508).