Large Language Models (LLMs) excel in natural language processing tasks but face deployment challenges due to high computational and memory demands. While quantization and parameter-efficient fine-tuning (PEFT) methods like LoRA offer solutions, low-bit quantization often leads to accuracy degradation and convergence issues. To address these, we propose GA-LoFTQ (Gradient-Aware Alternating Least Squares Framework for LoRA Fine-Tuning Quantization), a framework integrating Alternating Least Squares (ALS) optimization and gradient approximation to jointly optimize quantized weights and low-rank matrices. Extensive experiments on natural language understanding (NLU) and generation (NLG) tasks show GA-LoFTQ consistently outperforms methods like LoFTQ and QLoRA, especially under extreme low-bit settings (e.g., 2-bit). For example, GA-LoFTQ achieves 91.0% accuracy on QNLI, surpassing LoFTQ by 7%, with faster convergence and improved stability. These results demonstrate GA-LoFTQ’s ability to mitigate quantization-induced performance loss, enabling efficient and high-performance deployment of LLMs.

In ultrasound elastography (USE), B-mode (BM) data obtained from radiofrequency (RF) signals suffer significant information loss, limiting speckle tracking accuracy. Since RF data are typically inaccessible in clinical practice, reliable tracking from BM images is critical. We propose KDCLPWC-Net, an unsupervised optical flow network combining cross-modal knowledge distillation and curriculum learning to enhance displacement estimation from BM data. A parallelized Feature Alignment Module (FAM) is developed to transfer multi-scale RF features to the student network. Experiments on simulated and in vivo datasets show superior performance over existing methods, demonstrating KDCLPWC-Net’s potential for clinical elastography without requiring ground-truth labels.

Digital core reconstruction plays a vital role in subsurface analysis, but conventional techniques face challenges such as data scarcity, limited sample diversity, and the high cost of high-resolution CT imaging. Additionally, existing GAN-based models often suffer from geometric distortions and training instabilities, which compromise microstructural fidelity. To address these limitations, we propose a novel framework that integrates three key components: (1) StyleGAN3 for high-resolution image synthesis, (2) SIFT-based registration for precise structural alignment, and (3) a GAN-based inverse reconstruction pipeline for resolution enhancement. Evaluations on the Estaillades carbonate dataset demonstrate that our method achieves a 2.86-fold resolution improvement over conventional CT reconstruction, reducing reliance on high-end imaging equipment while preserving fine-scale microstructural details. This approach offers a scalable and cost-effective solution for digital rock analysis and reservoir simulation, advancing the state of the art in pore-scale modeling.

With the rapid advancement of computer vision and graphics, high-quality single-view 3D reconstruction has become increasingly significant in applications such as autonomous driving, robotics, and virtual reality. Traditional methods often struggle with the inherent ill-posed nature of single-view reconstruction, leading to limited accuracy and reduced detail retention. To address these challenges, we introduce an innovative framework that integrates an asymmetric U-Net architecture with 3D Gaussian splatting for single-view 3D reconstruction. Furthermore, our approach incorporates a novel 3D smoothing filter that constrains the maximum frequency of the 3D representation, effectively mitigating high-frequency artifacts in out-of-distribution rendering. By synergistically combining implicit and explicit representations, our method leverages the strengths of both to enhance reconstruction efficiency and quality. Experiments conducted on the SRN-Cars dataset demonstrate that our framework outperforms existing methods in quantitative metrics and qualitative assessments, achieving higher reconstruction accuracy and smoother surfaces.

Hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis (CA) are both heart diseases. Their echocardiographic presentations are similar, whereas the latter is rare in clinical practice, making them susceptible to misdiagnosis. Recently, deep learning-based echocardiographic diagnostic methods have been proposed. These data-driven methods require a large amount of data for training to improve diagnostic accuracy. Therefore, these methods do not apply to conditions with sparse clinical data, such as CA. To address the problem of poor model performance due to few training data, we propose a multimodal model (USCL) combining multiview and text to differentiate between HCM and CA. Specifically, the USCL consists of two main parts: one part employs unimodal supervision (UMS) to obtain echocardiogram and text respectively, the relevant unimodal representations and prediction results. The other part uses a designed multimodal contrast learning (MMCL) approach to tune the representations of the two modalities, thus exploring more reliable multimodal representations in few-shot learning. We build a dataset for training and evaluation which includes 209 patients with HCM and 203 patients with CA, as well as 200 patients with normal cardiac function. Experiments show that USCL achieves promising results.

This article presents a dual-branch Collaborative Siamese network architecture designed for visual object tracking, which we refer to as CoSiNet. The dual-branch collaborative Siamese network comprises two network branches: a shallow branch, which focuses on target localization to enhance resistance to interference from objects with similar characteristics, and a deep branch, which emphasizes the extraction of more abstract semantic information related to the object. Furthermore, we have devised a Channel Attention Feature Enhancement Module and a Spatial Channel Attention Feature Enhancement Module to augment feature extraction while mitigating the influence of background noise. In the concluding stages, an Adaptive Fusion Module is employed to amalgamate the response maps from both branches, resulting in an enhanced final response map. Experimental results, conducted on two publicly available datasets, demonstrate that our algorithm outperforms other state-of-the-art techniques in terms of tracking performance.

In ultrasound elastography, the noise present in the measured displacement fields has been a critical factor affecting the quality of the strain or elastic distribution reconstruction. Existing partial differential equation (PDE) based denoising algorithms can effectively remove noise to some extent but are limited by slow solving speeds. To address this challenge, we introduce a neural operator-based framework for denoising measured displacement fields in ultrasound elastography. By utilizing neural operators to learn a general solution operator for the denoising equation, the framework aims to achieve fast and accurate denoising of displacement fields. Experiments conducted on simulated data and tissue-mimicking phantom data demonstrate that this method performs comparably to traditional methods across various metrics. In particular, the proposed approach achieves processing speeds approximately 30 to 577 times faster than the finite difference method (FDM) in various data dimensions, making it a highly effective solution for rapid and scalable data processing in ultrasound elastography.

Ultrasound localization microscopy (ULM) is an innovative imaging technique that employs microbubbles (MBs) to improve the spatial resolution of ultrasound (US) imaging. Accurately extracting the MB signals from the original ultrasound data is essential for successful ULM. Traditional MB filtering methods, such as SVD, have high complexity and computational intensity. Due to the sensitivity to spatiotemporal information, 3D convolutional neural networks (3D CNN) have been utilized in MB filtering. However, the large number of parameters in 3D convolutional layers and complex network architectures affect the real-time performance of ULM. To optimize the network structure of 3D CNN and reduce parameters, this study proposes a novel MB filtering method based on 3D CNN and U-Net3+ named 3D U-Net3+. It adopts full-scale connection strategy to reduce network parameters, while combining low-level semantics and high-level semantics to capture fine-grained semantics and coarse-grained semantics at full scale. The experimental results demonstrate that the proposed MB filtering method can effectively preserve the spatiotemporal information of MBs in ultrasound sequence images. The SSIM and PSNR values of the MB image processed by the proposed method achieve 0.9141 and 30.881 dB, respectively. The obtained ULM image shows the microvessels as small as 20μm,

Coronary atherosclerotic disease is a major cause of myocardial infarction and usually causes partial or complete coronary artery disorders, which can be life-threatening in severe cases. Currently, the diagnosis of coronary atherosclerosis is mainly achieved through intravascular ultrasound (IVUS). By using intravascular ultrasound, the location and morphology of lesions can be identified early, which is crucial for the early detection and accurate diagnosis of coronary artery disease. Intravascular ultrasound is a widely used imaging technique for diagnosing and treating cardiovascular diseases. In this paper, we propose a new method for three-dimensional (3D) reconstruction of IVUS vascular models based on semantic segmentation. The proposed method utilizes state-of-the-art deep learning techniques to accurately segment the blood vessels in IVUS images. The proposed method utilizes state-of-the-art deep learning techniques to accurately segment the blood vessels in IVUS images. Subsequently, the segmented vessels are used to generate 3D reconstructions of the vascular models. Our method achieves high accuracy and robustness, and it has the potential to enhance the accuracy of IVUS-based diagnosis and treatment planning. Experimental results on a dataset of real-world IVUS images demonstrate the effectiveness of our method.

As human life continues improving, applications like virtual reality (VR) and augmented reality (AR) necessitate increasingly higher-quality 3D reconstructions. With the advancements in neural implicit 3D surface and volume rendering, multi-view 3D reconstruction has garnered significant attention. A prevailing issue in this domain is that the direct combination of neural implicit surface and volume rendering typically considers only photometric consistency loss, leading to an under-constrained surface problem. To address this problem, we develop a double-supervised approach and an integrated network for implicit surface and volume rendering, enabling the generation of a 3D surface model of an object from a set of multi-view images. In our approach, the object surface is represented by a signed distance function, and a 3D surface reconstruction model is jointly trained with photometric consistency constraints and geometry constraints. Experiments demonstrate that by enhancing prior geometry supervision and integrating the double-supervised network architecture into neural implicit 3D reconstruction, we can achieve accurate and high-quality 3D reconstruction.

In this paper, we propose a dual-branch Siamese network for visual object tracking. The proposed network consists of two distinct branches: a shallow network branch and a deep network branch. The shallow network branch focuses on precisely locating the target object and improving the anti-interference ability to similar objects, while the deep network branch focuses on capturing the more abstract semantic features of the object. Additionally, a multi-scale key feature fusion module is embedded into the shallow network, enabling the model to accurately locate the target object. Furthermore, we leverage the attention mechanism to further enhance the robustness of the model. Experimental results on three different public datasets demonstrate that our method outperforms state-of-the-art tracking algorithms.

Ultrasound localization microscopy (ULM) is a new imaging technique that uses microbubbles (MBs) to improve the spatial resolution of ultrasound (US) imaging. For ULM, it is critical to accurately localize MB position. Recently, deep learning-based methods are adopted to acquire MB localization, which shows promising performance and efficient computation. However, detection of high-concentration MBs is still a challenging task. To further improve the localization accuracy, a dual generative adversarial network (DualGAN)-based ULM imaging method (DualGAN-ULM) is proposed in this paper to overcome the problems of long data processing time and low parameter robustness in current ULM imaging methods. This method is trained using simulated data generated by point spread function (PSF) convolution and uses dual generation adversarial strategy to enable the generator to perform accurate localization under high-concentration MB conditions. Meanwhile, the localization and reconstruction capabilities of five ULM methods, namely Centroid, CS-ULM, mUNET-ULM, mSPCN-ULM and DualGAN-ULM, are evaluated in this paper. The experimental results reveal that DL-based ULM methods (DualGAN-ULM, mSPCN-ULM, and mUNET-ULM) outperform compressed sensing-based localization methods (CS-ULM) and Centroid in terms of localization accuracy and localization dependability. DualGAN-ULM performs better than mSPCN-ULM and mUNET-ULM, making it a more realistic ULM method.

Scene background initialization is an important step as one low-layer method for high-layer applications in computer vision. However, this process is always affected by practical challenges such as illumination changes, back-ground motion, camera jitter, intermittent movement and bad weather outdoors, etc. In this work, we develop a novel method called co-occurrence pixel-block (CPB) model via spatial-temporal correlation for robust back-ground initialization. This work first introduces the CPB method for foreground extraction. And then, background information in spatial-temporal features are utilized to recover an adaptive background for the current frame. Experimental results obtained from the dataset of the challenging benchmark (SBMnet) validate it’s performance under various challenges.

Accurately tracking large tissue motion over a sequence of ultrasound images is critically important to several clinical applications including, but not limited to, elastography, flow imaging, and ultrasound-guided motion compensation. However, tracking in vivo large tissue deformation in 3D is a challenging problem and requires further developments. In this study, we explore a novel tracking strategy that combines Bayesian inference with local polynomial fitting. Since this strategy is incorporated into a region-growing block-matching motion tracking framework we call this strategy a Bayesian region-growing motion tracking with local polynomial fitting (BRGMTLPF) algorithm. More specifically, unlike a conventional block-matching algorithm, we use a maximum posterior probability density function to determine the “correct” three-dimensional displacement vector.The proposed BRGMT-LPF algorithm was evaluated using a tissue-mimicking phantom and ultrasound data acquired from a pathologically-confirmed human breast tumor. The in vivo ultrasound data was acquired using a 3D whole breast ultrasound scanner, while the tissue-mimicking phantom was acquired using an experimental CMUT ultrasound transducer. To demonstrate the effectiveness of combining Bayesian inference with local Polynomial fitting, the proposed method was compared to the original region-growing motion tracking algorithm (RGMT), region-growing with Bayesian interference only (BRGMT), and region-growing with local polynomial fitting (RGMT-LPF). Our preliminary data demonstrate that the proposed BRGMT-LPF algorithm can improve the accuracy of motion tracking.

In recent years, plane wave (PW) ultrasound (US) imaging has emerged as a promising method for ultrafast US imaging due to its high temporal resolution. PW transmissions have a lower image quality when the number of PW is limited. The conventional approach to reconstruction entails a coherently summing sequence of US signals at the cost of frame rate. In this paper, an improved deep neural network approach is applied for reconstructing a high-quality PW image from a single PW radio frequency (RF) signal. Specifically, the deformable convolution is utilized as a feature extractor added on the U-net network, and a discriminator is utilized to improve the network fitting effect. It aims to learn a mapping between a single PW and compounding 75 PWs by training the network with in vitro and in vivo samples. The performance of the proposed approach is evaluated in terms of structural similarity (SSIM) and peak signal-to-noise ratio (PSNR). Extensive quantitative and visual evaluations reveal that the proposed model improves PSNR by 32.97 percent and 21.24 percent in cross-section and longitudinal sections of the carotid artery respectively, compared to the U-net. The results suggest the potential of reconstructing high-quality images from a single PW via deep learning.

Tissue motion tracking is a critically important step for many ultrasound elastography applications. In this study, we are particularly interested in evaluating motion tracking strategies for large deformation quasi-static elastography. In this study, Bayesian inference is incorporated into a region-growing motion estimation framework and we named the proposed tracking algorithm as a region-growing Bayesian motion tracking (RGBMT) algorithm. Basically, we replace signal correlation by a maximum posterior probability density function to perform motion tracking. Using a computer-simulated phantom and one set of human subject ultrasound data with pathologically-confirmed breast cancer, the proposed RGBMT algorithm was compared to the original region-growing motion tracking algorithm. Our preliminary data suggested that the addition of Bayesian inference is useful in terms of improving the accuracy of motion tracking. Results from both the numerical phantom and in vivo ultrasound data set showed that there are fewer tracking errors in axial displacement and strain images obtained from the proposed RGBMT algorithms. That explained why the contrast-to-noise (CNR) values were higher and the breast tumor on the reconstructed modulus image was better visualized.

Ultrasound elastography is a modality that is uniquely suited to augment conventional B-mode ultrasound for various clinical applications. Motion tracking plays a critically important role during image formation for ultrasound elastography. In this study, the accuracy of four motion tracking methods tailored for acoustic radiation force-based elastography (e.g. acoustic radiation force imaging, shear wave elastography) is compared. In these elastography methods, external mechanical excitation results in small tissue displacements (i.e. 5-10 micrometers). This paper compares four published motion tracking methods: a quadratic sub-sample estimation method, a coupled sub-sample estimation method, a 2-D spline-based estimator, and a 2-D autocorrelation-based motion estimator. Those four methods are evaluated using computer-simulated and tissue-mimicking phantom data. Based on our preliminary data, we find that the autocorrelation-based method is the preferred estimator without considering the lateral displacement. Overall, the spline-based estimator is superior to the other two competitors when both axial and lateral displacements are estimated. Since the spline-based estimation algorithm is considerably time-intensive, the coupled sub-sample estimation method becomes a practical alternative.

Monte-Carlo ray tracing, which enables realistic simulation of ultrasound-tissue interactions such as soft shadows and fuzzy reflections, has been used to simulate ultrasound images. The main technical challenge presented with Monte-Carlo ray tracing is its computational efficiency. In this study, we investigated the use of a commercially-available ray-tracing engine (NVIDIA's Optix 6.0), which provides a simple, recursive, and flexible pipeline for accelerating ray tracing algorithms. Our preliminary results show that our ultrasound simulation algorithm accelerated by the Optix engine can achieve a frame of 25 frames/second using an Nvidia RTX 2060 card. Furthermore, we compare ultrasound simulations built on the proposed Monte-Carlo ray-tracing algorithm with a deep-learning generative adversarial network (GANs)-based ultrasound simulator and a physics-based ultrasound simulator (Field II). The proposed ultrasound simulator was able to better visualize small-sized structures while the other two above-mentioned simulators could not. Our future work includes integration of our proposed simulator with a virtual reality platform and expansion to other ultrasound modalities such as elastography and flow imaging.

This paper presents an artificial intelligence-based ultrasound simulator suitable for medical simulation and clinical training. Particularly, we propose a machine learning approach to realistically simulate ultrasound images based on generative adversarial networks (GANs). Using B-mode ultrasound images simulated by a known ultrasound simulator, Field II, an "image-to-image" ultrasound simulator was trained. Then, through evaluations, we found that the GAN-based simulator can generate B-mode images following Rayleigh scattering. Our preliminary study demonstrated that ultrasound B-mode images from anatomies inferred from magnetic resonance imaging (MRI) data were feasible. While some image blurring was observed, ultrasound B- mode images obtained were both visually and quantitatively comparable to those obtained using the Field II simulator. It is also important to note that the GAN-based ultrasound simulator was computationally efficient and could achieve a frame rate of 15 frames/second using a regular laptop computer. In the future, the proposed GAN-based simulator will be used to synthesize more realistic looking ultrasound images with artifacts such as shadowing.

Advancements in medical ultrasound systems, challenges both manufacturers and clinicians in finding image functions and parameters that optimize image quality. We propose a machine learning method based on data mining for finding the best data path for certain exam types. Attribute relevance analysis helps us identify weakly relevant image parameters. The searching of frequent itemsets using apriori algorithm offers the best combination of image functions and their associated parameters. A commercially available ultrasound scanner was modified for our data collection, algorithmic verification, and analysis. Test results show that our proposed data mining methods may help manufacturers identify the most useful clinical image functions and help doctors choose right parameters as default settings that increase patient's throughput.