Development and validation of an AI algorithm to generate realistic and meaningful counterfactuals for retinal imaging based on diffusion models

PLOS Digital Health, Journal Year: 2025, Volume and Issue: 4(5), P. e0000853 - e0000853

Published: May 15, 2025

Counterfactual reasoning is often used by humans in clinical settings. For imaging based specialties such as ophthalmology, it would be beneficial to have an AI model that can create counterfactual images, illustrating answers questions like “If the subject had diabetic retinopathy, how fundus image looked?”. Such could aid training of clinicians or patient education through visuals answer queries. We large-scale retinal datasets containing color photography (CFP) and optical coherence tomography (OCT) images train ordinary adversarially robust classifiers classify healthy disease categories. In addition, we trained unconditional diffusion generate diverse including ones with lesions. During sampling, then combined classifier guidance achieve realistic meaningful maintaining subject’s structure. found our method generated counterfactuals introducing removing necessary disease-related features. conducted expert study validate are clinically meaningful. Generated were indistinguishable from real shown contain OCT appeared realistic, but identified experts higher than chance probability. This shows combining models even for high-resolution medical CFP images. professionals.

Language: Английский

---

Backdoor Training Paradigm in Generative Adversarial Networks

Entropy, Journal Year: 2025, Volume and Issue: 27(3), P. 283 - 283

Published: March 9, 2025

Backdoor attacks remain a critical area of focus in machine learning research, with one prominent approach being the introduction backdoor training injection mechanisms. These mechanisms embed triggers into process, enabling model to recognize specific trigger inputs and produce predefined outputs post-training. In this paper, we identify unifying pattern across existing methods generative models propose novel paradigm. This paradigm leverages unified loss function design facilitate diverse models. We demonstrate effectiveness generalizability through experiments on adversarial networks (GANs) Diffusion Models. Our experimental results GANs confirm that proposed method successfully embeds triggers, enhancing model’s security robustness. work provides new perspective methodological framework for models, making significant contribution toward improving safety reliability these

Language: Английский

---

Signal Super Prediction and Rock Burst Precursor Recognition Framework Based on Guided Diffusion Model with Transformer

Applied Sciences, Journal Year: 2025, Volume and Issue: 15(6), P. 3264 - 3264

Published: March 17, 2025

Implementing precise and advanced early warning systems for rock bursts is a crucial approach to maintaining safety during coal mining operations. At present, FEMR data play key role in monitoring providing warnings bursts. Nevertheless, conventional are associated with certain limitations, such as short time low accuracy of warning. To enhance the timeliness bolster mines, novel model has been developed. In this paper, we present framework predicting signal deep future recognizing burst precursor. The involves two models, guided diffusion transformer super prediction an auxiliary was applied Buertai database, which recognized having risk. results demonstrate that can predict 360 h (15 days) using only 12 known signal. If duration compressed by adjusting CWT window length, it becomes possible over longer spans. Additionally, achieved maximum recognition 98.07%, realizes disaster. These characteristics make our attractive

Language: Английский

---

Patrick Star: A comprehensive benchmark for multi-modal image editing

BenchCouncil Transactions on Benchmarks Standards and Evaluations, Journal Year: 2025, Volume and Issue: unknown, P. 100201 - 100201

Published: April 1, 2025

Language: Английский

---