Generative Modeling of the Circle of Willis Using 3D-StyleGAN DOI Creative Commons
Orhun Utku Aydin, Adam Hilbert, Alexander Koch

et al.

medRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: April 3, 2024

Abstract The circle of Willis (CoW) is a network cerebral arteries with significant inter-individual anatomical variations. Deep learning has been used to characterize and quantify the status CoW in various applications for diagnosis treatment cerebrovascular disease. In medical imaging, performance deep models limited by diversity size training datasets. To address data scarcity, generative adversarial networks (GANs) have applied generate synthetic vessel neuroimaging data. However, proposed methods produce fidelity or downstream utility tasks concerning characteristics. We adapted StyleGANv2 architecture 3D synthesize Time-of-Flight Magnetic Resonance Angiography (TOF MRA) volumes CoW. For modeling, we 1782 individual TOF MRA scans from 6 open source train StyleGAN model employed differentiable augmentations mixed precision cropped region interest 32×128×128 tackle computational constraints. was evaluated quantitatively using Fréchet Inception Distance (FID), MedicalNet distance (MD) Area Under Curve Precision Recall Distributions (AUC-PRD). Qualitative analysis performed via visual Turing test. demonstrated generated task multiclass semantic segmentation arteries. Vessel assessed Dice coefficient Hausdorff distance. best-performing high-quality diverse (FID: 12.17, MD: 0.00078, AUC-PRD: 0.9610). Multiclass trained on alone achieved comparable real most conclusion, modeling Circle synthesis paves way generalizable future, extensions provided methodology other imaging problems modalities inclusion pathological datasets potential advance development more robust clinical applications.

Language: Английский

Brain tumor segmentation using synthetic MR images - A comparison of GANs and diffusion models DOI Creative Commons
Muhammad Usman Akbar, Måns Larsson, Ida Blystad

et al.

Scientific Data, Journal Year: 2024, Volume and Issue: 11(1)

Published: Feb. 29, 2024

Abstract Large annotated datasets are required for training deep learning models, but in medical imaging data sharing is often complicated due to ethics, anonymization and protection legislation. Generative AI such as generative adversarial networks (GANs) diffusion can today produce very realistic synthetic images, potentially facilitate sharing. However, order share images it must first be demonstrated that they used different with acceptable performance. Here, we therefore comprehensively evaluate four GANs (progressive GAN, StyleGAN 1–3) a model the task of brain tumor segmentation (using two networks, U-Net Swin transformer). Our results show trained on reach Dice scores 80%–90% when real memorization problem models if original dataset too small. conclusion viable option further work required. The generated shared AIDA hub.

Language: Английский

Citations

22
Generative AI for synthetic data across multiple medical modalities: A systematic review of recent developments and challenges DOI Creative Commons
Mahmoud K. Ibrahim, Yasmina Al Khalil, Sina Amirrajab

et al.

Computers in Biology and Medicine, Journal Year: 2025, Volume and Issue: 189, P. 109834 - 109834

Published: March 1, 2025

This paper presents a comprehensive systematic review of generative models (GANs, VAEs, DMs, and LLMs) used to synthesize various medical data types, including imaging (dermoscopic, mammographic, ultrasound, CT, MRI, X-ray), text, time-series, tabular (EHR). Unlike previous narrowly focused reviews, our study encompasses broad array modalities explores models. Our aim is offer insights into their current future applications in research, particularly the context synthesis applications, generation techniques, evaluation methods, as well providing GitHub repository dynamic resource for ongoing collaboration innovation. search strategy queries databases such Scopus, PubMed, ArXiv, focusing on recent works from January 2021 November 2023, excluding reviews perspectives. period emphasizes advancements beyond GANs, which have been extensively covered reviews. The survey also aspect conditional generation, not similar work. Key contributions include broad, multi-modality scope that identifies cross-modality opportunities unavailable single-modality surveys. While core techniques are transferable, we find methods often lack sufficient integration patient-specific context, clinical knowledge, modality-specific requirements tailored unique characteristics data. Conditional leveraging textual conditioning multimodal remain underexplored but promising directions findings structured around three themes: (1) Synthesis highlighting clinically valid significant gaps using synthetic augmentation, validation evaluation; (2) Generation identifying personalization innovation; (3) Evaluation revealing absence standardized benchmarks, need large-scale validation, importance privacy-aware, relevant frameworks. These emphasize benchmarking comparative studies promote openness collaboration.

Language: Английский

Citations

5
Enhancing robustness and generalization in microbiological few-shot detection through synthetic data generation and contrastive learning DOI Creative Commons
Nikolas Ebert, Didier Stricker,

Oliver Wasenmüller

et al.

Computers in Biology and Medicine, Journal Year: 2025, Volume and Issue: 191, P. 110141 - 110141

Published: April 19, 2025

Language: Английский

Citations

0
On the Trustworthiness Landscape of State-of-the-art Generative Models: A Survey and Outlook DOI
Mingyuan Fan, Chengyu Wang, Cen Chen

et al.

International Journal of Computer Vision, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 28, 2025

Language: Английский

Citations

0
Beware of Diffusion Models for Synthesizing Medical Images - a Comparison with Gans in Terms of Memorizing Brain MRI and Chest X-Ray Images DOI
Muhammad Usman Akbar, Wuhao Wang, Anders Eklund

et al.

Published: Jan. 1, 2023

Diffusion models were initially developed for text-to-image generation and are now being utilized to generate high quality synthetic images. Preceded by GANs, diffusion have shown impressive results using various evaluation metrics. However, commonly used metrics such as FID IS not suitable determining whether simply reproducing the training Here we train StyleGAN models, BRATS20, BRATS21 a chest x-ray pneumonia dataset, synthesize brain MRI images, measure correlation between images all Our show that more likely memorize compared StyleGAN, especially small datasets when 2D slices from 3D volumes. Researchers should be careful medical imaging, if final goal is share

Language: Английский

Citations

4
Effect of Training Epoch Number on Patient Data Memorization in Unconditional Latent Diffusion Models DOI
Salman Ul Hassan Dar,

Isabelle Ayx,

Marie Kapusta

et al.

Informatik aktuell, Journal Year: 2024, Volume and Issue: unknown, P. 88 - 93

Published: Jan. 1, 2024

Language: Английский

Citations

0
Artificial intelligence in cardiovascular imaging and intervention DOI

Sandy Engelhardt,

Salman Ul Hussan Dar,

Lalith Sharan

et al.

Herz, Journal Year: 2024, Volume and Issue: 49(5), P. 327 - 334

Published: Aug. 9, 2024

Language: Английский

Citations

0
Conditional 4D Motion Diffusion Models with Masked Observations to Forecast Deformations DOI

Sylvain Thibeault,

Liset Vázquez Romaguera, Samuel Kadoury

et al.

Lecture notes in computer science, Journal Year: 2024, Volume and Issue: unknown, P. 89 - 98

Published: Jan. 1, 2024

Language: Английский

Citations

0
Latent Pollution Model: The Hidden Carbon Footprint in 3D Image Synthesis DOI

Marvin Seyfarth,

Salman Ul Hassan Dar, Sandy Engelhardt

et al.

Lecture notes in computer science, Journal Year: 2024, Volume and Issue: unknown, P. 146 - 156

Published: Oct. 5, 2024

Language: Английский

Citations

0
Generative Modeling of the Circle of Willis Using 3D-StyleGAN DOI Creative Commons
Orhun Utku Aydin, Adam Hilbert, Alexander Koch

et al.

medRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: April 3, 2024

Abstract The circle of Willis (CoW) is a network cerebral arteries with significant inter-individual anatomical variations. Deep learning has been used to characterize and quantify the status CoW in various applications for diagnosis treatment cerebrovascular disease. In medical imaging, performance deep models limited by diversity size training datasets. To address data scarcity, generative adversarial networks (GANs) have applied generate synthetic vessel neuroimaging data. However, proposed methods produce fidelity or downstream utility tasks concerning characteristics. We adapted StyleGANv2 architecture 3D synthesize Time-of-Flight Magnetic Resonance Angiography (TOF MRA) volumes CoW. For modeling, we 1782 individual TOF MRA scans from 6 open source train StyleGAN model employed differentiable augmentations mixed precision cropped region interest 32×128×128 tackle computational constraints. was evaluated quantitatively using Fréchet Inception Distance (FID), MedicalNet distance (MD) Area Under Curve Precision Recall Distributions (AUC-PRD). Qualitative analysis performed via visual Turing test. demonstrated generated task multiclass semantic segmentation arteries. Vessel assessed Dice coefficient Hausdorff distance. best-performing high-quality diverse (FID: 12.17, MD: 0.00078, AUC-PRD: 0.9610). Multiclass trained on alone achieved comparable real most conclusion, modeling Circle synthesis paves way generalizable future, extensions provided methodology other imaging problems modalities inclusion pathological datasets potential advance development more robust clinical applications.

Language: Английский

Citations

0