CT Neural Network Loss Functions for Frequency-Preserving Image Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network training methods for medical imaging, such as CT, often fail to align subjective radiologist preferences with objective image quality, requiring extensive training and are inflexible to changes in imaging protocols, leading to oversmoothing and loss of high frequencies.
Innovation Solution
Tailored neural networks using novel loss functions that penalize specific error components, such as frequency and spatial features, to optimize image properties, allowing for flexible training and preservation of desired image characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If conventional pointwise loss functions are used for training neural networks, then training is straightforward and computationally efficient, but the resulting images fail to align with radiologist preferences and exhibit oversmoothing
Solution Approach 1:
The patent transforms the conventional pointwise loss function into a frequency-domain loss function by applying Fourier transforms and defining loss in terms of frequency components. This parameter transformation allows the network to preserve important frequency information while maintaining training efficiency, thereby improving image quality alignment without sacrificing operational simplicity
Solution Approach 2:
The patent introduces frequency domain representation as an intermediary between the input and output domains. By defining the loss function in the frequency domain and using Fourier transforms as intermediaries, the system bridges the gap between computational efficiency and radiologist preference alignment, allowing gradient computation to flow through the frequency domain without requiring direct manipulation of complex image data
2Manufacturing precision
If extensive training on large image ensembles is performed, then image quality may improve, but training time and computational resources increase significantly
Solution Approach 1:
By changing the domain from spatial to frequency domain for loss computation, the patent enables more effective training with fewer samples. The frequency domain loss function captures essential image characteristics more efficiently, allowing the network to learn from smaller ensembles while achieving comparable or superior image quality to extensive spatial domain training
3Ease of manufacture
If conventional training methods are used, then the network can be trained once, but it requires entirely new training sequences when CT protocols or imaging technology change
Solution Approach 1:
The frequency domain loss function serves multiple purposes: it works across different CT protocols, imaging technologies, and corruption types. By defining the loss in terms of frequency components rather than protocol-specific parameters, the same training approach can be universally applied to various imaging scenarios, reducing the need for protocol-specific retraining
Solution Approach 2:
The patent enables dynamic adaptability by allowing the frequency domain loss function to be adjusted for different protocols and technologies. The frequency-based approach provides a flexible framework where the same fundamental training method can adapt to changing conditions without requiring complete retraining sequences
4Manufacturing precision
If the network is trained to reduce noise and artifacts, then image quality improves, but important diagnostic features may be removed along with the noise
Solution Approach 1:
By transforming the loss function to the frequency domain, the patent enables selective processing of different frequency components. The network can be trained to remove noise in specific frequency ranges while preserving diagnostic features in other frequency ranges, as the frequency domain loss function allows for frequency-selective optimization that maintains important signal components
Data Source
AI summary
Techniques are described that facilitate generating neural network (NNs) tailored to optimize specific properties of medical images using novel loss functions. According to an embodiment, a system is provided that comprises a memory that stores computer executable components, and a processor that executes the computer executable components stored in the memory. The computer executable components comprise a training component that trains a NN to generate a modified version of computed tomography (CT) data comprising one or more optimized properties relative to the CT data using a loss function tailored to control learning adaptation of the NN based on error attributed to one or more defined components associated with the CT data, resulting in a trained NN, wherein the one or more defined components comprise at least one of a frequency component or a spatial feature component.