Dehazing Ultrasound using Diffusion Models

IEEE Transactions on Medical Imaging, Journal Year: 2024, Volume and Issue: 43(10), P. 3546 - 3558

Published: Feb. 7, 2024

Echocardiography has been a prominent tool for the diagnosis of cardiac disease. However, these diagnoses can be heavily impeded by poor image quality. Acoustic clutter emerges due to multipath reflections imposed layers skin, subcutaneous fat, and intercostal muscle between transducer heart. As result, haze other noise artifacts pose real challenge ultrasound imaging. In many cases, especially with difficult-to-image patients such as obesity, from B-Mode imaging is effectively rendered unusable, forcing sonographers resort contrast-enhanced examinations or refer modalities. Tissue harmonic popular approach combat haze, but in severe cases still impacted haze. Alternatively, denoising algorithms are typically unable remove highly structured correlated noise, It remains accurately describe statistical properties develop an inference method subsequently it. Diffusion models have emerged powerful generative shown their effectiveness variety inverse problems. this work, we present joint posterior sampling framework that combines two separate diffusion model distribution both clean unsupervised manner. Furthermore, demonstrate techniques training on radio-frequency data highlight advantages over data. Experiments in-vitro in-vivo datasets show proposed dehazing removes while preserving signals weakly reflected tissue.

Language: Английский

---

Generative artificial intelligence: synthetic datasets in dentistry

BDJ Open, Journal Year: 2024, Volume and Issue: 10(1)

Published: March 1, 2024

Abstract Introduction Artificial Intelligence (AI) algorithms, particularly Deep Learning (DL) models are known to be data intensive. This has increased the demand for digital in all domains of healthcare, including dentistry. The main hindrance progress AI is access diverse datasets which train DL ensuring optimal performance, comparable subject experts. However, administration these traditionally acquired challenging due privacy regulations and extensive manual annotation required by Biases such as ethical, socioeconomic class imbalances also incorporated during curation datasets, limiting their overall generalizability. These challenges prevent accrual at a larger scale training models. Methods Generative techniques can useful production Synthetic Datasets (SDs) that overcome issues affecting datasets. Variational autoencoders, generative adversarial networks diffusion have been used generate SDs. following text review operations. It discusses chances SDs with potential solutions will improve understanding healthcare professionals working research. Conclusion customized need researchers produced robust models, having trained on dataset applicable dissemination across countries. there limitations associated better understood, attempts made those concerns prior widespread use.

Language: Английский

---

A review of self‐supervised, generative, and few‐shot deep learning methods for data‐limited magnetic resonance imaging segmentation

NMR in Biomedicine, Journal Year: 2024, Volume and Issue: 37(8)

Published: March 24, 2024

Abstract Magnetic resonance imaging (MRI) is a ubiquitous medical technology with applications in disease diagnostics, intervention, and treatment planning. Accurate MRI segmentation critical for diagnosing abnormalities, monitoring diseases, deciding on course of treatment. With the advent advanced deep learning frameworks, fully automated accurate advancing. Traditional supervised techniques have tremendously, reaching clinical‐level accuracy field segmentation. However, these algorithms still require large amount annotated data, which oftentimes unavailable or impractical. One way to circumvent this issue utilize that exploit limited labeled data. This paper aims review such state‐of‐the‐art use number samples. We explain fundamental principles self‐supervised learning, generative models, few‐shot semi‐supervised summarize their cardiac, abdomen, brain Throughout review, we highlight can be employed based quantity data available. also present comprehensive list notable publicly available datasets. To conclude, discuss possible future directions field—including emerging algorithms, as contrastive language‐image pretraining, potential combinations across methods discussed—that further increase efficacy image labels.

Language: Английский

---

DF-CDM: Conditional diffusion model with data fusion for structural dynamic response reconstruction

Mechanical Systems and Signal Processing, Journal Year: 2024, Volume and Issue: 222, P. 111783 - 111783

Published: Aug. 1, 2024

Language: Английский

---

Investigating Data Memorization in 3D Latent Diffusion Models for Medical Image Synthesis

Lecture notes in computer science, Journal Year: 2024, Volume and Issue: unknown, P. 56 - 65

Published: Jan. 1, 2024

Language: Английский

---