Research

Mobile ultrasound devices play a crucial role in emergency and family doctor services, providing valuable diagnostic capabilities in a portable form factor. However, when compared to traditional medical ultrasound equipment, there is still scope for improvement in terms of image quality.

This study has introduced some innovations to enhance the super-resolution reconstruction of medical ultrasound images, particularly for mobile ultrasound devices. By optimizing the CycleGAN network’s generator with a novel set of hyperparameters and convolution kernels, our method has successfully extracted multiscale features more effectively, thereby significantly improving the quality of the reconstructed images and making them controllable. Additionally, employing paired and unpaired ultrasound images alongside natural images in the training process has enabled our model to acquire a more robust understanding of ultrasound-specific features. This is an advancement in leveraging the capabilities of natural cognition to enhance medical imaging. Moreover, the perceptual loss module addresses the critical issue of information loss that plagues traditional methods. This module’s ability to deeply extract and integrate features into the network loss function marks a leap forward in pursuing perceptual consistency and fidelity in ultrasound image super-resolution.

Ultrasound imaging is widely used in clinical diagnosis due to its advantages such as safety, non-invasiveness, convenience, and ease of operation. However, actual ultrasound images often suffer from problems such as speckle noise, low signal-to-noise ratio, and low contrast, requiring denoising of the ultrasound images.

During the process of ultrasound image acquisition, due to the reflection, scattering, and refraction characteristics of ultrasound echoes, as well as the heterogeneity and spatially uncertain properties of different parts of the human body, a large number of randomly distributed scattering particles are formed when ultrasound waves penetrate the human body. The interactions between scattering particles generate correlated scattering beams. During the process of echo reflection, interference effects from the interference of reflected echoes and mutual interference between scattering beams cause amplitude enhancement and attenuation due to the different phases of the echoes when different beams overlap. This leads to random fluctuations in the electrical signals in the output after envelope detection by the transducer, and generates speckle particles with different brightness levels in ultrasound images. This type of noise is commonly referred to as multiplicative speckle noise. Additionally, during the operation of the ultrasound device, internal components, circuits, electromagnetic interference, etc., produce additive Gaussian noise. These types of noise collectively contribute to the complex noise in medical ultrasound images.

Currently, researchers have studied denoising methods for medical ultrasound images. However, these methods cannot adaptively denoise the noise generated by different settings of medical ultrasound devices.

To overcome the challenges encountered in denoising ultrasound images, we propose a controllable method for denoising medical ultrasound images. This study mainly utilizes the TATLAB Toolbox to generate simulated noisy ultrasound image datasets and constructs an initial denoising model using the preprocessing of simulated noisy ultrasound images, the multi-level residual atrous spatial pyramid pooling (MRASPP) module, the nonlinear mapping convolutional neural network (NMCNNB) module, and the adaptive noise level and variable denoising intensity module. The model is trained through two rounds of inference.

This method can adapt to a wider range of denoising scenarios, adaptively denoise the images, flexibly adjust the denoising intensity, effectively remove noise from medical ultrasound images, and preserve the details of the images. Compared to existing denoising methods, the average values of SSIM and PSNR have been improved by 1.67% and 1.28%, respectively. A comparison was made between the denoised and original images to evaluate their performance in downstream image classification tasks. After denoising real breast ultrasound images, the ACC and AUC of tumor benign-malignant classification tasks were improved by 1.09% and 2.83%, respectively.

The proposed controllable method for denoising medical ultrasound images, using two rounds of inference, can adaptively select the noise level parameter for effective denoising of ultrasound images formed by different ultrasound devices and settings. This method provides a new approach for denoising other types of medical images.

Please note that the translation and revision have been done to the best of my abilities, but there might still be room for improvement.

Magnetic resonance (MR) images are crucial for diagnosis, but noise and artifacts can reduce image quality and affect diagnostic accuracy. Previous studies have mostly dealt with denoising or artifact removal separately, ignoring pixel correlation and noise level changes, relying on manual parameter adjustment, and lacking uncertainty evaluation, which limits practical applications. To address these issues, we propose an Adaptive Dual-branch Network for Denoising and Artifact Removal in MR Images (ADbNet) that includes two branches: denoising and artifact removal. The heterogeneous window transformer branch is used for remote and short-range feature modeling and achieves adaptive denoising while preserving complex structures and textures. The improved UNet branch is used to remove artifacts from MR images. The last two branches are integrated through an attention based fusion module to improve image quality. In the inference stage, calculate uncertainty helps doctors determine the reliability of the results. Our experimental results on the Brats2021 and OpenNeuro datasets show that the PSNR, SSIM and NMSE values of ADbNet are 27.23 ± 1.02 dB, 0.9132±0.0134 and 0.0753±0.0093, 27.13 ± 1.32 dB, 0.9121 ± 0.0458 and 0.0765±0.0091, respectively. Compared with state-of-the-art methods, ADbNet exhibits excellent performance in denoising and artifact removal tasks. The improved image joint uncertainty map was applied to downstream task medical image segmentation. The segmentation results showed that the average Dice values of MSD brain dataset T1n, T1c, T2f, and T2w modalities increased by 2.27%, 1.33%, 1.37%, and 1.04%, respectively. For the T2w and Adc sequences of the Prostate dataset, the average Dice values increased by 1.11% and 1.18%, respectively, indicating an improvement in segmentation performance.

Image Quality Improvement of Mobile Ultrasound

A Transducer-adaptive Denoising Model for Medical Ultrasound Imaging

ADbNet: Adaptive Dual-branch Network for Denoising and Artifact Removal in MR Images