RadImageGAN – A Multi-modal Dataset-Scale Generative AI for Medical Imaging

Lecture notes in computer science, Год журнала: 2025, Номер unknown, С. 173 - 185

Опубликована: Янв. 1, 2025

---

Synthetic tabular data generation in Federated Learning environments: A practical use case for Acute Myeloid Leukemia (Preprint)

Опубликована: Март 20, 2025

Data scarcity and dispersion pose significant obstacles in biomedical research, particularly when addressing rare diseases. In such scenarios, Synthetic Generation (SDG) has emerged as a promising path to mitigate the first issue. Concurrently, Federated Learning (FL) is machine learning paradigm where multiple nodes collaborate create centralized model with knowledge that distilled from data different nodes, but without need for sharing it. This research explores combination of SDG FL technologies context Acute Myeloid Leukemia, hematological disorder, evaluating their combined impact quality generated artificial datasets. To evaluate privacy- fidelity-related federating distribution scenarios numbers comparing them baseline model. A state-of-the-art Generative Adversarial Network architecture was trained considering four scenarios: (1) non-federated all available, (2) federated scenario evenly distributed among (3) unevenly randomly (imbalanced data), (4) non-IID distributions. For each fixed set node quantities (3, 5, 7, 10) considered assess its impact, evaluated attending fidelity-privacy trade-off. The computed fidelity metrics exhibited statistically deteriorations (P < 0.001) ranging 0.21% 21.23% due federation process. When experiments diverse no strong tendencies were observed, even if specific comparisons resulted significative differences. Privacy mainly maintained while obtaining maximum improvements 55.17% 26.23, although they not significant. Within scope use case this paper, act an algorithm results loss compared maintaining privacy levels. However, deterioration does significantly increase number used train models grows, though differences found comparisons. fact amount differently neither most nor metrics, similar scenarios.

Язык: Английский

---

Privacy enhancing and generalizable deep learning with synthetic data for mediastinal neoplasm diagnosis

npj Digital Medicine, Год журнала: 2024, Номер 7(1)

Опубликована: Окт. 20, 2024

Язык: Английский

---

Synthetic Image Generation of Aortic Valves Using Conditional DDPM

Опубликована: Сен. 13, 2024

Язык: Английский

---

Optimization research and analysis of the basic pathway of diffusion model based on big data algorithm

Опубликована: Июль 19, 2024

Язык: Английский

---

Generative Artificial Intelligence Approaches for Synthesizing High-Fidelity Breast Thermal Images

Lecture notes in computer science, Год журнала: 2024, Номер unknown, С. 33 - 43

Опубликована: Ноя. 2, 2024

Язык: Английский

---

Synthetic dual image generation for reduction of labeling efforts in semantic segmentation of micrographs with a customized metric function

Methods in microscopy, Год журнала: 2024, Номер unknown

Опубликована: Дек. 13, 2024

Abstract Training of semantic segmentation models for material analysis requires micrographs as the inputs and their corresponding masks. In this scenario, it is quite unlikely that perfect masks will be drawn, especially at edges objects, sometimes amount data can obtained small, since only a few samples are available. These aspects make very problematic to train robust model. Therefore, we demonstrate in work an easy-to-apply workflow improvement through generation synthetic microstructural images conjunction with The joining respective create input Vector Quantised-Variational AutoEncoder (VQ-VAE) model includes embedding space, which trained such generative (PixelCNN) learns distribution each input, transformed into discrete codes, used sample new codes. latter eventually decoded by VQ-VAE generate alongside segmentation. To evaluate quality generated data, have U-Net different amounts these real data. were then evaluated using microscopic only. Additionally, introduce customized metric derived from mean Intersection over Union (mIoU) excludes classes not part ground-truth mask when calculating mIoU all classes. proposed prevents falsely predicted pixels greatly reducing value mIoU. With implemented workflow, able achieve time reduction preparation acquisition, well image processing labeling tasks. idea behind approach could generalized various types serves user-friendly solution training smaller number images.

Язык: Английский

---