CT Neural Network Self-Training for Noise Reduction Without Large Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep convolutional neural networks (CNNs) for improving CT image quality require large amounts of high-quality and low-quality data for training, which may not always be available, limiting their applicability and generalizability.
Innovation Solution
A self-training method is employed to generate noise-reduced medical images by using data augmentation techniques such as random rotation and noise addition directly on projection data from a single patient, generating paired low-quality and high-quality images for training a neural network without relying on external data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep CNN-based methods are used to improve CT image quality, then image quality is improved, but large amounts of high-quality and low-quality training data are required which may not always be available
Solution Approach 1:
The system performs self-training by generating synthetic training data from its own processed images. The trained CNN model processes CT images, and the output images are then used as training data to further refine the model, creating a self-improving loop that eliminates dependency on external large-scale training datasets
Solution Approach 2:
The system pre-processes CT images through the trained CNN model before formal training begins. These pre-processed images serve as preliminary training data that are then used to initialize and guide the subsequent training process, allowing the model to start with some prior knowledge rather than from scratch
2Reliability
If conventional CNNs are trained on external datasets, then training performance is achieved, but generalizability to new patients and data distributions is limited
Solution Approach 1:
The system dynamically adapts training parameters including learning rate, batch size, and training epochs based on the specific characteristics of each patient's data. This allows the model to optimize its training process for each individual case, improving generalizability across different patients and data distributions
Solution Approach 2:
The training process is made dynamic and adaptive rather than static. The model continuously adjusts its parameters and architecture based on feedback from processing each patient's data, enabling it to adapt to different data distributions and maintain high performance across diverse clinical scenarios
Data Source
AI summary
Noise-reduced images of a subject are generated from x-ray projection data acquired from the subject using a computed tomography (“CT”) system. A neural network or other machine learning algorithm is trained to receive images reconstructed from the projection data as an input and to generate an output as noise-reduced images. The neural network or other machine learning algorithm is trained using a self-training procedure, in which the training data used to train the neural network or other machine learning algorithm are generated directly from the projection data acquired from the subject using data augmentation (e.g., random rotations of the projection data and/or noise insertion).