Ilias Sarbout

Ilias Sarbout

Auto-generated profile, based on publicly available information

Ilias Sarbout

Affiliation Sorbonne Université; CNRS; Inserm; Inria; Paris Brain Institute – ICM; AP-HP, Paris, France

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 July 2025 in Journal of the American Heart Association
Verified
Journal Article
Open access
AGAyse Gungor
Ayse Gungor
ISIlias Sarbout
Ilias Sarbout
AGAubrey L. Gilbert
Aubrey L. Gilbert
SHSteffen Hamann
Steffen Hamann
PLPierre Lebranchu
Pierre Lebranchu
DMDan Milea
Dan Milea
Background: Prompt diagnosis of acute central retinal artery occlusion (CRAO) is crucial for therapeutic management and stroke prevention. However, most stroke centers lack onsite ophthalmic expertise before considering fibrinolytic treatment. This study aimed to develop, train, and test a deep learning system to detect hyperacute CRAO on retinal fundus photographs within the critical 4.5‐hour treatment window and up to 24 hours after visual loss to aid in secondary stroke prevention. Methods: Our retrospective, cross‐sectional study included 1322 color fundus photographs from 771 patients with acute visual loss due to CRAO, central retinal vein occlusion, nonarteritic anterior ischemic optic neuropathy, and healthy controls. Photographs were collected from 9 expert neuro‐ophthalmology centers in 6 countries, including 3 randomized clinical trials. Training included 1039 photographs (517 patients), followed by testing on 2 data sets: (1) hyperacute CRAO (54 photographs, 54 patients) and (2) CRAO within 24 hours after visual loss (110 photographs, 109 patients). Results: The deep learning system achieved an area under the receiver operating characteristic curve of 0.96 (95% confidence interval (CI), 0.95–0.98), a sensitivity of 92.6% (95% CI, 87.0–98.0), and a specificity of 85.0% (95% CI, 81.8–92.8) for detecting CRAO at hyperacute stage, with similar results within 24 hours. The deep learning system outperformed stroke neurologists on a subset of hyperacute testing data set (120 photographs, 120 patients). Conclusions: A deep learning system can accurately detect hyperacute CRAO on retinal photographs within a time window compatible with urgent fibrinolysis. If further validated, such systems could improve patient selection for fibrinolytic trials and optimize secondary stroke prevention. Registration: URL: https://www.clinicaltrials.gov ; Unique identifier: NCT06390579.

21 October 2025 in PLOS Computational Biology
Verified
Journal Article
Open access
PDF file
MOMehdi Ounissi
Mehdi Ounissi
ISIlias Sarbout
Ilias Sarbout
JHJean-Pierre Hugot
Jean-Pierre Hugot
CMChristine Martinez-Vinson
Christine Martinez-Vinson
DBDominique Berrebi
Dominique Berrebi
DRDaniel Racoceanu
Daniel Racoceanu
Chemical staining methods, while reliable, are time consuming and can be resource-intensive, involving costly chemical reagents and raising environmental concerns. This underscores the compelling need for alternative solutions such as virtual staining, which not only accelerates the diagnostic process but also enhances the flexibility of stain applications without the associated physical and chemical costs. Generative artificial intelligence technologies prove to be immensely useful in addressing these challenges. However, in healthcare, particularly within computational pathology, the high-stakes nature of decisions complicates the adoption of these tools due to their often opaque processes. Our work introduces an innovative approach that harnesses generative models for virtual stain transformations, improving performance, trustworthiness, scalability, and adaptability within computational pathology. The core of the proposed methodology involves a singular Hematoxylin and Eosin (H&E) encoder that supports multiple stain decoders. This design prioritizes critical regions in the latent space of H&E tissues, leading to a richer representation that enables precise synthetic stain generation by the decoders. Tested to simultaneously generate eight different stains from a single H&E slide, our method also offers significant scalability benefits for routine use by loading only necessary model components during production. We integrate label-free knowledge during training, using loss functions and regularization to minimize artifacts, thereby enhancing the accuracy of virtual staining in both paired and unpaired settings. To build trust in these synthetic stains, we employ a real-time self-inspection methodology using trained discriminators for each stain type, providing pathologists with confidence heatmaps to aid in their evaluations. In addition, we perform automatic quality checks on new H&E slides to ensure that they conform to the trained H&E distribution, guaranteeing the generation of high-quality synthetic stained slides. Recognizing the challenges pathologists face in adopting new technologies, we have encapsulated our method in an open-source, cloud-based proof-of-concept system. This system enables users to easily and virtually stain their H&E slides through a browser, eliminating the need for specialized technical knowledge and addressing common hardware and software challenges. It also facilitates real-time user feedback integration. Lastly, we have curated a novel dataset comprising eight different paired H&E/stains related to pediatric Crohn’s disease at diagnosis, providing 30 whole slide images (WSIs) for each stain set (total of 480 WSIs) to stimulate further research in computational pathology.

5 May 2026 in Array
Verified
Journal Article
Open access
ISIlias Sarbout
Ilias Sarbout
MOMehdi Ounissi
Mehdi Ounissi
DMDan Milea
Dan Milea
DRDaniel Racoceanu
Daniel Racoceanu