SELFNet: Denoising Shear Wave Elastography Using Spatial-temporal Fourier Feature Networks.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Deep Learning

Background:

• Ultrasound shear wave elastography (SWE) estimates tissue stiffness by analyzing shear wave propagation.
• Limited acquisitions in SWE can introduce noise into displacement or velocity fields.
• Physics-informed deep learning offers a novel approach to approximate physics fields by minimizing governing equation residuals.

Purpose of the Study:

• To introduce the shear wave elastography Fourier feature network (SELFNet) for estimating and denoising particle displacement signals.
• To leverage spatial-temporal random Fourier features within a physics-informed neural network (PINN) framework.
• To simultaneously learn shear modulus mapping and incorporate governing equations for regularization.

Main Methods:

• Developed SELFNet using a physics-informed deep learning framework.
• Incorporated spatial-temporal random Fourier features and sparse mapping for robustness.
• Evaluated the network on tissue-mimicking phantoms and ex vivo tissue datasets.

Main Results:

• SELFNet effectively smoothed noise in phantom lesions of varying stiffness and sizes.
• Achieved superior performance over Gaussian filtering, with 17% lower relative ℓ₂ error and 45% lower RMSE.
• Ablation studies confirmed that the Fourier feature mapping module prevents overfitting.

Conclusions:

• SELFNet demonstrates significant potential for improving SWE with limited acquisitions.
• The method's applicability was confirmed in ex vivo tissue studies.
• Successful translation could lead to clinical applications in diagnosing diseases like cancer and liver disease.