Deep Sylvester Posterior Inference for Adaptive Compressed Sensing in Ultrasound Imaging

Published: Jan. 13, 2025

Ultrasound images are commonly formed by sequential acquisition of beam-steered scan-lines. Minimizing the number required scan-lines can significantly enhance frame rate, field view, energy efficiency, and data transfer speeds. Existing approaches typically use static subsampling schemes in combination with sparsity-based or, more recently, deep-learning-based recovery. In this work, we introduce an _adaptive_ method that maximizes intrinsic information gain _in-situ_, employing a Sylvester Normalizing Flow encoder to infer approximate Bayesian posterior under partial observation real-time. Using deep generative model for future observations, determine scheme mutual between subsampled next video. We evaluate our approach using EchoNet cardiac ultrasound video dataset demonstrate active sampling outperforms competitive baselines, including uniform variable-density random sampling, as well equidistantly spaced scan-lines, improving mean absolute reconstruction error 15%. Moreover, inference generation performed just 0.015 seconds (66Hz), making it fast enough real-time 2D imaging applications.

Language: Английский

---

Comparative analysis of dehazing algorithms on real-world hazy images

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: March 28, 2025

Abstract Images captured in adverse weather conditions (haze, fog, smog, mist, etc.) often suffer significant degradation. Due to the scattering and absorption of these particles, various negative effects, such as reduced visibility, low contrast, colour distortion are introduced into image. These degraded images unsuitable for many computer vision applications, including smart transportation, video surveillance, forecasting, remote sensing. To ensure reliable operation a high-quality haze-free input image is essential, which supplied by dehazing techniques. This review categorises recent methods, highlighting popular approaches within each group. In years, deep learning methods restoration-based techniques using priors have garnered attention, particularly addressing challenges dense non-homogeneous haze. this paper, their typical candidates compared real-world hazy because most data-driven neural augmentation trained synthetic images. Experimental results conducted on reveal that physics-driven single-image algorithms exhibit lack robustness, while perform well thin but struggle haze conditions. Neural algorithms, however, effectively combine strengths both approaches, offering better overall solution. By identifying existing gaps paper provides valuable resource novice experienced researchers, pointing towards future directions rapidly advancing field.

Language: Английский

---

Deep Learning-Based Medical Ultrasound Image and Video Segmentation Methods: Overview, Frontiers, and Challenges

Sensors, Journal Year: 2025, Volume and Issue: 25(8), P. 2361 - 2361

Published: April 8, 2025

The intricate imaging structures, artifacts, and noise present in ultrasound images videos pose significant challenges for accurate segmentation. Deep learning has recently emerged as a prominent field, playing crucial role medical image processing. This paper reviews video segmentation methods based on deep techniques, summarizing the latest developments this such diffusion segment anything models well classical methods. These are classified into four main categories characteristics of Each category is outlined evaluated corresponding section. We provide comprehensive overview learning-based methods, evaluation metrics, common datasets, hoping to explain advantages disadvantages each method, summarize its achievements, discuss future trends.

Language: Английский

---

Enhancing Prediction of Heart Disease Using Hybrid Machine Learning Methods XGBoost and K-Means Clustering

Research Square (Research Square), Journal Year: 2025, Volume and Issue: unknown

Published: April 17, 2025

Efficient Heart Disease Prediction by Hybrid Machine Learning Methods disease is a major killer around the world and has much promise using machine learning model techniques highly problematic concerning privacy risk as well lack handling of heterogeneous (non-IID) data institutions. This paper aims to create experiment with hybrid framework integrating Extreme Gradient Boosting (XGBoost) K-Means Clustering in enhancing predictive power bringing together powerful ensemble unsupervised pattern detection. The was tested trained on benchmark databases such UCI Cleveland datasets had centralized architecture optimized for strength interpretability Performance baseline models Logistic Regression, Naive Bayes, Random Forest accuracy, precision, recall, F1-score, AUC-ROC, computational efficiency XGBoost-KMeans performed best 93.8% 92.4% 93.1% 92.7% AUC-ROC 0.96, outperforming 5–6% accuracy generalization classifying patient groups comparable factors. KMeans module enabled enhanced feature representation clustering, while XGBoost high-accuracy classification. Overall, system provides robust explainable solution predict heart performance healthcare analytics systems deployable scalability clinical settings IoT wearable sources.

Language: Английский

---

Distributed Learning for Heart Disease Risk Prediction Based on Key Clinical Parameters with Evaluation Metrics Analysis

Research Square (Research Square), Journal Year: 2025, Volume and Issue: unknown

Published: April 21, 2025

The purpose of this study design and test a Decentralized Federated learning framework that integrates Mutual Learning approach with Hierarchical Dirichlet Process-based (HDP-FLDec) algorithm to predict heart disease. Conventional machine models in health care are data-centralized, which poses major risks patient privacy lacks support for handling data heterogeneity between institutions. Current federated also hard-pressed by non-IID lack scalability. Privacy protecting prediction disease is early treatment intervention. A decentralized intelligent method can improve diagnostic accuracy without compromising data, kept local secure.A framework, combining mutual clients employing HDP as model the latent structure diverse data.The endpoints were correctness AUC-ROC. Secondary precision, recall, F1-score, communication efficiency network.Heart sets from public benchmark sources, namely UCI Heart Disease Cleveland datasets varying demographic clinical records.Performance measures calculated over several training peer-to-peer topology against centralized standard algorithms (FedAvg, FedProx).The HDP-FLDec attained better performance all baseline models.The scored an 94.2%, precision 92.8%, recall 93.6%, F1-score 93.2%, AUC-ROC 0.97.HDP-FLDec beat 5–7% showed improved stability under conditions.The mechanism speed convergence, while aspect delivered personalization generalization. Communication was reduced 18% compared architectures.The suggested provides privacy-preserving accurate environment, efficiently dealing limitations.The offers practical solution healthcare systems real life enhance simultaneously preserving confidentiality patients. Integration real-time monitoring IoT-based medical devicesHDP-FLDec marries Bayesian nonparametrics' flexibility decentralization offered presenting greater strength, scalability existing models.

Language: Английский

---

Score-based Diffusion Models for Photoacoustic Tomography Image Reconstruction

ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Journal Year: 2024, Volume and Issue: unknown, P. 2470 - 2474

Published: March 18, 2024

Photoacoustic tomography (PAT) is a rapidly-evolving medical imaging modality that combines optical absorption contrast with ultrasound depth. One challenge in PAT image reconstruction inadequate acoustic signals due to limited sensor coverage or the density of transducer array. Such cases call for solving an ill-posed inverse problem. In this work, we use score-based diffusion models solve problem reconstructing from measurements. The proposed approach allows us incorporate expressive prior learned by model on simulated vessel structures while still being robust varying sparsity conditions.

Language: Английский

---

Diffusion Models for Medical Image Reconstruction

Deleted Journal, Journal Year: 2024, Volume and Issue: 1(1)

Published: Jan. 1, 2024

Abstract Better algorithms for medical image reconstruction can improve quality and enable reductions in acquisition time radiation dose. A prior understanding of the distribution plausible images is key to realising these benefits. Recently, research into deep-learning has started look using unsupervised diffusion models, trained only on high-quality (ie, without needing paired scanner measurement data), modelling this understanding. Image incorporating models have already attained state-of-the-art accuracy tasks ranging from highly accelerated MRI ultra-sparse-view CT low-dose PET. Key advantages model approach over previous deep learning approaches include modelling, improved robustness domain shift, principled quantification uncertainty. If hallucination concerns be alleviated, their impressive performance could mean are better suited clinical use than approaches. In review, we provide an accessible introduction outline guidance diffusion-model-based methodology, summarise modality-specific challenges, identify themes. We conclude with a discussion opportunities challenges reconstruction.

Language: Английский

---

TransDehaze: transformer-enhanced texture attention for end-to-end single image dehaze

The Visual Computer, Journal Year: 2024, Volume and Issue: unknown

Published: May 27, 2024

Language: Английский

---

Denoising Plane Wave Ultrasound Images Using Diffusion Probabilistic Models

IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control, Journal Year: 2024, Volume and Issue: 71(11), P. 1526 - 1539

Published: Aug. 26, 2024

Ultrasound plane wave imaging is a cutting-edge technique that enables high frame-rate imaging. However, one challenge associated with ultrasound the noise them, hindering their wider adoption. Therefore, development of denoising method becomes imperative to augment quality images. Drawing inspiration from Denoising Diffusion Probabilistic Models (DDPMs), our proposed solution aims enhance image quality. Specifically, considers distinction between low-angle and high-angle compounding waves as effectively eliminates it by adapting DDPM beamformed radiofrequency (RF) data. The underwent training using only 400 simulated In addition, approach employs natural segmentation masks intensity maps for generated images, resulting in accurate various anatomy shapes. was assessed across simulation, phantom, vivo results evaluations indicate not enhances on data but also demonstrates effectiveness phantom terms Comparative analysis other methods underscores superiority evaluation metrics. source code trained model will be released along dataset at: http://code.sonography.ai.

Language: Английский

---

Synthesizing Real-Time Ultrasound Images of Muscle Based on Biomechanical Simulation and Conditional Diffusion Network

IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control, Journal Year: 2024, Volume and Issue: 71(11), P. 1501 - 1513

Published: Aug. 26, 2024

Quantitative muscle function analysis based on the ultrasound imaging, has been used for various applications, particularly with recent development of deep learning methods. The nature speckle noises in images poses challenges to accurate and reliable data annotation supervised algorithms. To obtain a large dataset without manual scanning labelling, we proposed synthesizing pipeline provide synthetic datasets movement an ground truth, allowing augmenting, training, evaluating models different tasks. Our contained biomechanical simulation using finite element method, algorithm reconstructing sparse fascicles, diffusion network image generation. With adjustment few parameters, can generate real-time diversity morphology pattern. 3,030 generated, qualitatively quantitatively verified that closely matched in-vivo images. In addition, applied into tasks analysis. Compared trained unaugmented dataset, model one had better cross-dataset performance, which demonstrates feasibility augment training avoid over-fitting. results regression task show potentials under conditions number or label are limited. not only be muscle-related study, but other similar study development, where sequential needed training.

Language: Английский

---