Enhancing Radiographic Diagnosis: CycleGAN-Based Methods for Reducing Cast Shadow Artifacts in Wrist Radiographs

Deleted Journal, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 3, 2025

We extend existing techniques by using generative adversarial network (GAN) models to reduce the appearance of cast shadows in radiographs across various age groups. retrospectively collected 11,500 adult and paediatric wrist radiographs, evenly divided between those with without casts. The test subset consisted 750 cast. extended results from a previous study that employed CycleGAN enhancing model perceptual loss function self-attention layer. which incorporates layer delivered similar quantitative performance as original model. This was applied images 20 cases where reports recommended CT scanning or repeat cast, were then evaluated radiologists for qualitative assessment. demonstrated generated could improve radiologists' diagnostic confidence, some leading more decisive reports. Where available, follow-up imaging compared produced reading AI-generated images. Every report, except two, provided identical diagnoses associated imaging. ability perform robust reporting downsampled AI-enhanced is clinically meaningful warrants further investigation. Additionally, unable distinguish unenhanced These findings suggest suppression technique be integrated tool augment clinical workflows, potential benefits reducing patient doses, improving operational efficiencies, delays diagnoses, number visits.

Language: Английский

---

Application of CycleGAN-based image style transfer algorithm in visual communication design

Journal of Computational Methods in Sciences and Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 10, 2025

This paper addresses the challenges in traditional image style transfer research, including high design costs, reliance on paired data, limited effects, and a lack of optimization for practical applications. To overcome these limitations, we introduce CycleGAN algorithm efficient without aiming to enhance efficiency, personalization, diversity required visual communication design. Using pre-trained model, conduct experiments with datasets obtained from DesignNet Google Open Images, following comprehensive preprocessing workflow. The effectiveness CycleGAN-based is quantified using Structural Similarity Index (SSIM). Experimental results demonstrate that model performs excellently transfer, achieving SSIM values 0.85, 0.84, 0.87 classical painting, modern art, retro poster styles, respectively. Further validation through user surveys expert evaluations confirms quality generated images terms aesthetics, clarity, recognizability, creativity score 4.5 satisfaction rate 90%. study not only enriches technical toolkit but also provides valuable insights into application potential CycleGAN, particularly tasks do require samples, showcasing algorithm’s unique advantages applicability.

Language: Английский

---

Coevolutionary generative adversarial networks for medical image augumentation at scale

Proceedings of the Genetic and Evolutionary Computation Conference, Journal Year: 2022, Volume and Issue: unknown

Published: July 8, 2022

Medical image processing can lack images for diagnosis. Generative Adversarial Networks (GANs) provide a method to train generative models data augmentation. Synthesized be used improve the robustness of computer-aided diagnosis systems. However, GANs are difficult due unstable training dynamics that may arise during learning process, e.g., mode collapse and vanishing gradients. This paper focuses on Lipizzaner, GAN framework combines spatial coevolution with gradient-based learning, which has been mitigate pathologies. Lipizzaner improves performance by taking advantage its distributed nature running at scale. Thus, algorithm implementation scaled high-performance computing (HPC) systems more accurate models. We address medical imaging augmentation create chest X-Ray using HPC infrastructure provided Oak Ridge National Labs' Summit Supercomputer. The experimental analysis shows improved increasing scale training. also demonstrate coevolutionary even when suboptimal neural network architectures hardware constraints.

Language: Английский

---

Unsupervised contrastive unpaired image generation approach for improving tuberculosis screening using chest X-ray images

Pattern Recognition Letters, Journal Year: 2022, Volume and Issue: 164, P. 60 - 66

Published: Oct. 27, 2022

Tuberculosis is an infectious disease that mainly affects the lung tissues. Therefore, chest X-ray imaging can be very useful to diagnose and understand evolution of pathology. This image modality has a poorer quality in contrast with other techniques as magnetic resonance or computerized tomography, but easier cheaper perform. Furthermore, data scarcity challenging domain biomedical imaging. In order mitigate this problem, use Generative Adversarial Network models for generation proved powerful approach train deep learning small datasets, representing alternative classic augmentation strategies. work, we propose fully automatic novel synthetic images effect improve tuberculosis screening performance using 3 different publicly available representative datasets: Montgomery County, Shenzhen TBX11K. Firstly, trains translation large-sized dataset (TBX11K). Then, these are used generate set small-sized medium-sized datasets (Montgomery County Shenzhen, respectively). Finally, generated added training screening. As result, obtained 88.41% ± 5.27% accuracy 90.33% 1.41% dataset. These results demonstrate proposed method outperforms previous state-of-the-art approaches.

Language: Английский

---

ShipGAN: Generative Adversarial Network based simulation-to-real image translation for ships

Applied Ocean Research, Journal Year: 2023, Volume and Issue: 131, P. 103456 - 103456

Published: Jan. 11, 2023

Recent advances in robotics and autonomous systems (RAS) have significantly improved the autonomy level of unmanned surface vehicles (USVs) made them capable undertaking demanding tasks various environments. During operation USVs, apart from normal situations, it is those unexpected scenes, such as busy waterways or navigation dust/nighttime, impose most dangers to USVs these scenes are rarely seen during training. Such a rare occurrence also makes manual collection recording into dataset difficult, expensive inefficient, with majority existing public available datasets not able fully cover them. One many plausible solutions purposely generate data using computer vision techniques assistance high-fidelity simulations that can create desirable motions/scenarios. However, stylistic difference between simulation images natural would cause domain shift problem. Hence, there need for designing method transfer distribution styles realistic domain. This paper proposes evaluates novel solution fill this gap Generative Adversarial Network (GAN) based model, ShipGAN, translate images. Experiments were carried out investigate feasibility generating GAN-based image translation models. The synthetic demonstrated be reliable by object detection segmentation algorithms trained

Language: Английский

---