Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PHOCUS: Physics-Based Deconvolution for Ultrasound Resolution Enhancement

Felix Duelmer, Walter Simson, Mohammad Farid Azampour, Magdalena Wysocki, Angelos Karlas, Nassir Navab

TL;DR

PHOCUS addresses ultrasound resolution limits by formulating image formation as a physics-based convolution and learning a continuous echogenicity map from B-mode images using Implicit Neural Representations. It introduces a differentiable rendering pipeline with a multi-resolution hash-encoded input to an INR, jointly optimizing the echogenicity map and PSF-informed blurring through a loss blending SSIM, L2, and TV terms. Compared to Richardson-Lucy on synthetic data, PHOCUS delivers higher PSNR/SSIM and sharper edges, and yields qualitative resolution gains on wire phantoms and in vivo carotid imaging, demonstrating practical applicability from common B-mode data. The work suggests future enhancements via learnable PSF models and broader clinical validation to further improve diagnostic ultrasound performance.

Abstract

Ultrasound is widely used in medical diagnostics allowing for accessible and powerful imaging but suffers from resolution limitations due to diffraction and the finite aperture of the imaging system, which restricts diagnostic use. The impulse function of an ultrasound imaging system is called the point spread function (PSF), which is convolved with the spatial distribution of reflectors in the image formation process. Recovering high-resolution reflector distributions by removing image distortions induced by the convolution process improves image clarity and detail. Conventionally, deconvolution techniques attempt to rectify the imaging system's dependent PSF, working directly on the radio-frequency (RF) data. However, RF data is often not readily accessible. Therefore, we introduce a physics-based deconvolution process using a modeled PSF, working directly on the more commonly available B-mode images. By leveraging Implicit Neural Representations (INRs), we learn a continuous mapping from spatial locations to their respective echogenicity values, effectively compensating for the discretized image space. Our contribution consists of a novel methodology for retrieving a continuous echogenicity map directly from a B-mode image through a differentiable physics-based rendering pipeline for ultrasound resolution enhancement. We qualitatively and quantitatively evaluate our approach on synthetic data, demonstrating improvements over traditional methods in metrics such as PSNR and SSIM. Furthermore, we show qualitative enhancements on an ultrasound phantom and an in-vivo acquisition of a carotid artery.

PHOCUS: Physics-Based Deconvolution for Ultrasound Resolution Enhancement

TL;DR

PHOCUS addresses ultrasound resolution limits by formulating image formation as a physics-based convolution and learning a continuous echogenicity map from B-mode images using Implicit Neural Representations. It introduces a differentiable rendering pipeline with a multi-resolution hash-encoded input to an INR, jointly optimizing the echogenicity map and PSF-informed blurring through a loss blending SSIM, L2, and TV terms. Compared to Richardson-Lucy on synthetic data, PHOCUS delivers higher PSNR/SSIM and sharper edges, and yields qualitative resolution gains on wire phantoms and in vivo carotid imaging, demonstrating practical applicability from common B-mode data. The work suggests future enhancements via learnable PSF models and broader clinical validation to further improve diagnostic ultrasound performance.

Abstract

Ultrasound is widely used in medical diagnostics allowing for accessible and powerful imaging but suffers from resolution limitations due to diffraction and the finite aperture of the imaging system, which restricts diagnostic use. The impulse function of an ultrasound imaging system is called the point spread function (PSF), which is convolved with the spatial distribution of reflectors in the image formation process. Recovering high-resolution reflector distributions by removing image distortions induced by the convolution process improves image clarity and detail. Conventionally, deconvolution techniques attempt to rectify the imaging system's dependent PSF, working directly on the radio-frequency (RF) data. However, RF data is often not readily accessible. Therefore, we introduce a physics-based deconvolution process using a modeled PSF, working directly on the more commonly available B-mode images. By leveraging Implicit Neural Representations (INRs), we learn a continuous mapping from spatial locations to their respective echogenicity values, effectively compensating for the discretized image space. Our contribution consists of a novel methodology for retrieving a continuous echogenicity map directly from a B-mode image through a differentiable physics-based rendering pipeline for ultrasound resolution enhancement. We qualitatively and quantitatively evaluate our approach on synthetic data, demonstrating improvements over traditional methods in metrics such as PSNR and SSIM. Furthermore, we show qualitative enhancements on an ultrasound phantom and an in-vivo acquisition of a carotid artery.
Paper Structure (14 sections, 10 equations, 4 figures, 1 table)

This paper contains 14 sections, 10 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Image Formation Process
  4. Proposed Framework
  5. Richardson-Lucy Algorithm
  6. Data
  7. In Silico Data:
  8. In-Vitro and In-Vivo:
  9. Experiments & Results
  10. Implementation Details
  11. Synthetic Dataset
  12. Wire Target In-Vitro
  13. Carotid Data In-Vivo
  14. Discussion & Conclusion

Figures (4)

  • Figure 1: General outline of the proposed framework:(1) encoding of the input location based on a multi-resolution look-up table, which forms the input to the MLP (inspired by mullerInstantNeuralGraphics2022). The neural network predicts the echogenicity map. (2) convolution of the echogenicity map with the PSF. (3) computation of the loss and backpropagation to update the MLP.
  • Figure 2: The left column represents the axial and lateral resolution targets, and the right column displays the cylindrical inclusions.
  • Figure 3: Predicted echogenicity map (center) of the B-mode image (left). The right side shows the result of the minimum enclosing circles' algorithm on the filtered and clustered image.
  • Figure 4: In-Vivo carotid data: the acquired B-mode image with 8 MHz central frequency, the predicted echogenicity map, and the reconstructed B-Mode image at the same frequency, at 6 MHz, and at 10 MHz