Machine Learning Spike Noise Detection in MRI K-Space Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Magnetic Resonance Imaging (MRI) scans suffer from spike noise, also known as popcorn or burst noise, which is unpredictable and randomly occurs due to static charge in metallic components, causing severe artifacts in diagnostic images that hinder clinical interpretation.
Innovation Solution
An image processing apparatus and method using machine learning classifiers to identify spike noise in k-space data by analyzing intensity values and distance from the origin, allowing for precise classification and correction of noise before image transformation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If noise removal algorithms are used to remove spike noise, then noise reduction is achieved, but the algorithms cannot be successfully deployed due to the unpredictable and random nature of spike noise
Solution Approach 1:
The patent transforms the classification problem from depending on unpredictable spike characteristics to depending on predictable parameters: intensity value and distance from k-space origin. By changing the basis of classification from temporal/spectral features to spatial-intensity features, the algorithm becomes reliable and deployable despite the random nature of spike noise occurrence
Solution Approach 2:
The patent replaces traditional signal processing approaches with a machine learning classification system. Instead of using conventional noise filtering mechanics, it employs a trained classifier that automatically identifies and removes spike noise based on learned patterns from training data, achieving both reliability and effectiveness
2Measurement precision
If machine learning classification is used to identify spike noise, then noise detection accuracy is improved, but system complexity increases
Solution Approach 1:
The patent segments the spike noise detection problem into distinct classification categories (spike noise vs. non-spike data samples). By dividing the k-space data into manageable classification units based on intensity and position, the system achieves high detection accuracy while keeping the processing complexity structured and controllable
Solution Approach 2:
The patent performs preliminary classification of k-space data samples before image reconstruction. By pre-identifying and flagging spike noise samples in the raw data domain using the trained classifier, the system achieves accurate noise detection without adding complexity to the subsequent reconstruction process, as the classification results are simply incorporated into the existing pipeline
3Manufacturing precision
If spike noise is removed from k-space data, then image quality is improved, but processing time increases
Solution Approach 1:
The patent applies partial action by classifying only the necessary k-space data samples for spike noise detection using the simple intensity-and-distance criteria. By not over-processing the data and focusing classification efforts only where needed, the system improves image quality while minimizing additional processing time
Solution Approach 2:
The patent uses a trained classifier model that was previously trained on training data. This copied knowledge from the training phase allows the system to quickly identify spike noise during actual processing without requiring complex real-time analysis, thus improving image quality while keeping processing time acceptable
Data Source
AI summary
An apparatus and method of detecting a characteristic in an image is performed by obtaining, from an image capturing apparatus, raw signal data formed from a plurality of data samples and including a signal of interest captured by the image capturing apparatus and classifying, using a neural network, samples other than the signal of interest using a classifier having been determined using a first parameter based on information about the sample and a second parameter based on information identifying a position of the sample within the raw image data.