Dental Image Denoising via Domain Adaptation and Synthetic Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Dental images, particularly radiographic scans, often lack true images due to high radiation doses and sensor blurring, making it challenging to train machine learning algorithms for effective denoising without introducing errors.
Innovation Solution
Utilizing domain adaptation and generated noise to train a neural network on one type of dental image (e.g., photos) and apply it to another (e.g., radiographic images) by creating data pairs with generated noise, allowing the network to learn a more accurate noise model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional denoising techniques are used on dental images, then noise is reduced, but small structures that truly belong to the underlying object are accidentally removed
Solution Approach 1:
The patent creates synthetic training data by copying and transforming real dental images, adding various types of noise to generate artificial noisy-image/true-image pairs. This allows the neural network to learn denoising patterns without needing actual ground truth radiographic images, thereby preserving small structures while removing noise.
Solution Approach 2:
The patent transforms images between different domains (photograph domain to radiographic domain) by changing parameters such as adding specific noise patterns, adjusting contrast, and simulating sensor characteristics. This domain transformation enables the network to learn denoising specific to radiographic images while preserving diagnostic structures.
2Measurement precision
If ground truth radiographic images are obtained to train machine learning algorithms, then training accuracy is improved, but radiation dose increases and sensor blurring occurs
Solution Approach 1:
The patent introduces photographs as an intermediary domain that can serve as ground truth references. By training the neural network to transform noisy radiographic images to match the corresponding photograph (which serves as the true image), the system achieves accurate training without requiring ground truth radiographic images, thus avoiding additional radiation exposure and sensor blurring.
Solution Approach 2:
The patent creates synthetic radiographic images by copying and transforming photographs with added noise and artifacts. This allows the generation of unlimited training data pairs without additional patient exposure, as the underlying true images (photographs) can be reused and transformed multiple times.
3Adaptability or versatility
If domain adaptation is used to train on one type of dental image and apply to another, then training data availability is improved, but model accuracy may deteriorate
Solution Approach 1:
The patent systematically transforms images between domains by applying specific parameter changes: adding radiographic noise patterns, adjusting contrast ratios, simulating sensor response characteristics, and modifying frequency content. These controlled parameter changes create realistic domain transitions while preserving the essential structural information needed for accurate denoising.
Solution Approach 2:
The patent creates synthetic training data by copying real photographs and transforming them into radiographic-like images with added noise and artifacts. This copying and transformation process generates large amounts of labeled training data without requiring additional clinical images, thereby improving data availability while maintaining denoising accuracy through the use of true photograph references.
Data Source
Figure 1
Figure 2
Figure 3A~3C
AI summary
Disclosed is a computer-implemented method of denoising dental images, comprising: obtaining a training set of images of a first type, comprising noisy images and true images, where the noisy image is the true image with generated noise; training at least one neural network on the training set; and denoising a dental image of the second type through the trained neural network.