Deep Learning Harmonic Imaging Training via Frequency-Weighted Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based harmonic imaging methods result in large and computationally intensive networks, limiting their applicability in real-world use due to memory and computational costs, while conventional methods suffer from artifacts, reduced contrast-to-noise ratio, and limited penetration depth.
Innovation Solution
A training method for deep learning-based harmonic imaging that reduces model size by leveraging frequency-based components, using a combination of filters to determine errors and update neural network parameters, emphasizing high-frequency components during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the complexity of the deep learning network is increased to enhance accuracy, then image quality improves, but memory and computational costs increase
Solution Approach 1:
The patent changes the parameter of the loss function by incorporating frequency-weighted filtering that emphasizes high-frequency components. This allows the network to achieve better image quality by focusing training on critical frequency regions rather than increasing network depth and channels, thus resolving the contradiction between image quality and model complexity
Solution Approach 2:
The patent applies local quality by differentiating the treatment of different frequency components in the loss function. High-frequency components are weighted more heavily to improve critical image features, while low-frequency components are treated differently. This targeted approach improves image quality without requiring a uniformly complex network structure throughout
2Measurement precision
If the complexity of the deep learning network is increased to enhance accuracy, then image quality improves, but computational power requirements increase
Solution Approach 1:
The patent modifies the loss function parameters to include frequency-based weighting that prioritizes high-frequency components. This parameter change enables the network to achieve superior image quality with a simpler architecture, thereby reducing computational power requirements while maintaining or improving performance
3Device complexity
If conventional harmonic imaging methods are used, then computational costs are reduced, but image quality deteriorates due to artifacts and reduced contrast-to-noise ratio
Solution Approach 1:
The patent replaces conventional signal processing methods with a deep learning-based approach. The neural network learns to reconstruct high-quality harmonic images directly from the data, substituting traditional filtering and processing techniques. This substitution improves image quality by eliminating artifacts and enhancing contrast-to-noise ratio while keeping computational costs manageable through efficient network design
Data Source
AI summary
An apparatus for training a model to perform harmonic imaging using ultrasound signals, the apparatus including processing circuitry configured to input first ultrasound data into a neural network model configured to generate and output second ultrasound data, determine a first error by applying a first filter to a difference between the second ultrasound data and target ultrasound data, determine a second error by applying a second filter, different from the first filter, to the difference between the second ultrasound data and the target ultrasound data, determine a loss value based on the determined first error and the determined second error, and update parameters of the neural network model based on the determined loss value to generate a trained neural network model.