Deep Learning Medical Image Denoising via Multi-Contrast Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image denoising methods, including deep learning techniques, face challenges in adapting to medical images with limited datasets and risk introducing artifacts or missing pathology details, particularly in improving the signal-to-noise ratio (SNR) and resolution of arterial spin labeling (ASL) MRI images, which require repeated scans increasing testing time.
Innovation Solution
A method utilizing multi-contrast information and deep learning to generate improved medical images by training a deep network model with data augmentation techniques such as cropping, rotating, and flipping, and incorporating non-local mean filtering to adaptively tune model parameters, thereby enhancing image quality without the need for repeated scans.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If repeated ASL scans are performed to increase SNR, then image quality is improved, but testing time is significantly increased
Solution Approach 1:
The patent applies preliminary action by using deep learning models trained on high-SNR reference images to predict and restore the low-SNR ASL images before actual scanning. The model learns the mapping from low-SNR to high-SNR image characteristics during training, allowing rapid inference during clinical use without requiring repeated scans.
Solution Approach 2:
The patent uses copying by creating synthetic high-SNR reference images from low-SNR ASL images through the deep learning model. The model copies the structural and diagnostic information from the low-SNR input and generates a synthetic high-SNR output, effectively replicating the quality of multiple repeated scans from a single scan.
2Reliability
If deep learning models are trained with large datasets, then model performance is improved, but data availability is insufficient for medical imaging
Solution Approach 1:
The patent applies self-service by using the deep learning model to generate its own training data. The model takes low-SNR ASL images, processes them through the network, and uses the output to create synthetic high-SNR reference images for training. This self-generated data expands the effective training dataset without requiring external data collection.
Solution Approach 2:
The patent uses disposable synthetic data generated from a single low-SNR scan to train the model. Instead of requiring expensive, time-consuming acquisition of multiple high-SNR reference scans for training, the system generates sufficient training data synthetically from the available low-SNR images, making the training process more efficient and scalable.
3Productivity
If conventional denoising methods are used, then processing speed is maintained, but image quality and pathology detection are compromised
Solution Approach 1:
The patent replaces conventional mechanical denoising algorithms (such as Gaussian filtering, wavelet filtering, and non-local means) with a deep learning-based approach. The neural network model learns complex non-linear mappings from noisy to clean images, substituting traditional signal processing mechanisms with data-driven learning that achieves superior image quality while maintaining processing speed through optimized inference.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A method of improving diagnostic and functional imaging, comprising obtaining at least one input image of a subject, using a medical imager, wherein said at least one input image comprises a first contrast, generating one or more copies of said at least one input image of said subject using spatial filtering, using an appropriately programmed computer, wherein said one or more copies comprises different spatial characteristics, applying, using said appropriately programmed computer, a trained deep network on said at least one input image and said one or more copies of said at least one input image of said subject to generate an output image, wherein said trained deep network comprises data augmentation and is trained based on at least one reference image, wherein said at least one reference image comprises a second contrast that is higher than said first contrast, and outputting the output image for analysis or visualization by a user..