Artificial Intelligence In Ophthalmology Center

Mahtab Faraji

Mahtab is currently pursuing her Ph.D. in Biomedical Engineering at the University of Illinois at Chicago (UIC). After completing a Bachelor’s degree in Biomedical Engineering from Sahand University of Technology, Iran, in 2014, she completed her Master’s degree in the same field at Iran University of Science and Technology, in 2018. Her research passion lies in applying Artificial Intelligence (AI) methods to enhance computer-aided diagnosis, with a primary emphasis on Ophthalmology and Cardiovascular medicine. In her free time, she enjoys social activities, running, and dancing.

Georgie Nahass

Georgie is a 5th year MSTP student in the AI-O center at UIC advised by Darvin Yi. His PhD thesis is on deep learning for periorbital scene understanding, with a specific focus on periorbital distance extraction and downstream clinical use cases in out of distribution settings. In his PhD he has developed numerous novel datasets specific to this task and has demonstrated that periorbital distances are robust features for disease classification and vision prediction. He is currently working on deploying his algorithms at clinics both in Chicago and around the world to expand the dataset utility and move towards large scale multimodal modeling of oculoplastic and craniofacial diseases. When not doing research, Georgie can be found playing guitar with his friends or in the gym trying to be less bad at basketball and/or running.
He received his undergraduate degrees in Molecular Biology and Biochemistry from Colorado College where he studied novel inhibitors of amyloid formation in neurodegeneration. Prior to medical school, he researched the mucosal antibody response to infectious diseases at Stanford University and MIT. Georgie chose to pursue a PhD in the AI-O lab as he believes AI can be leveraged to reduce disparities and inefficiencies in the healthcare system and increase both global and local access to ophthalmic care.

Virginia Tasso

Virginia is currently pursuing a Master’s degree in Biomedical Engineering through the joint program between Politecnico di Milano and the University of Illinois at Chicago. Her research interests are about computer vision applied to medical image analysis, with a particular interest in developing imaging-based solutions to advance personalized medicine, enabling patient-tailored therapeutic strategies and improving clinical decision-making. In her free time, she enjoys reading and swimming.

Alumni

• Homa Rashidisabet, PhD
• Abhishek Sethi, PhD
• Simon (Jiechao) Ma, MS
• Jacob Love, MS
• Minhaj Nur Alam, PhD
• Abdullah Aleem, MS
• Sasha Kravets, PhD
• Nooshin Mojab, PhD
• Kate Romond

George R. Nahass, Zhu Wang, Homa Rashidisabet, Won Hwa Kim, Sasha Hubschman, Jeffrey C. Peterson, Ghasem Yazdanpanah, Chad A. Purnell, Pete Setabutr, Ann Q. Tran, Darvin Yi, Sathya N. Ravi

In this project we introduce a novel formulation for targeted machine unlearning (or the removal of specific subsets of data from a deep learning network after training) using Langevin dynamics. We validated this algorithm on standard benchmark datasets, open source and clinically captured CFP images, and oculoplastic images collected the UIC ophthalmology clinic.

https://doi.org/10.48550/arXiv.2505.21872

In this project, we propose and validate a novel pipeline for periorbital distance prediction. We also demonstrate that AI predicted periorbital distances can be used as features to train shallow models for disease classification in a manner that is significantly more robust to OOD samples compared to conventional classification techniques.

George R. Nahass, Emma Koehler, Nicholas Tomaras, Danny Lopez, Madison Cheung, Alexander Palacios, Jeffrey C. Peterson, Sasha Hubschman, Kelsey Green, Chad A. Purnell, Pete Setabutr, Ann Q. Tran, Darvin Yi

Here, we created the first open-source large scale dataset for detailed periorbital segmentation. We annotated and validated 2800 images across two open facial recognition source datasets for the purpose of predicting the small, detailed anatomy around the eyes.

https://doi.org/10.1016/j.xops.2025.100757

In this paper, we present a tool for curating massive datasets collected over many years of clinical practice for the purpose of downstream training tasks related to oculoplastic and craniofacial surgery.

Driven by a need to deploy our algorithms in clinics with limited access to ophthalmic care, we designed and built a web application to serve our periorbital distance prediction models (Glorbit-global periorbital). Glorbit is translated into multiple languages and was beta tested under approved IRB. It is currently in use at Quina Care clinic in the Putamayo Canton of Ecuador.

Uveal Melanoma (UM), the most common intraocular cancer, is difficult to diagnose accurately due to its similarity to benign lesions and the limited access to ocular oncology experts. While deep learning (DL) has shown strong results across ophthalmology, its application to UM remains underexplored, primarily because of data scarcity. This work presents a benchmark study comparing classification, detection, and segmentation models for UM diagnosis under varying data regimes. Building on recent proposals to reframe dense-prediction tasks as classification, we extend these strategies to a real clinical setting, establishing robust baselines for automated UM diagnosis and providing practical guidance for model selection and adaptation in data-scarce environments.

I-ODA is a live databank that includes real-world longitudinal clinical ophthalmic images and patient metadata, with the purpose of advancing state-of-the-art computer vision applications in ophthalmology, and improving upon the translatable capacity of machine learning based applications.

Project lead: Joelle Hallak

Team: Nooshin Mojab, Abdullah Aleem, Manoj Prabhakar Nallabothula, Joe Baker, Darvin Yi

Auto Ptosis is a pipeline for detecting and calculating the features to help diagnosis ptosis. This disease presents itself with the “drooping” or “falling” of the upper eyelid and can lead to an impairment of visual field. We tackle this task through several different levels of hand-crafted design: (1) we train a classification network to differentiate healthy and ptosis eyes in an end-to-end fashion, (2) we use keypoint detection networks to delineate eyelids and pupils to derive the MRD1 distance (the marginal reflex distance-1, the gold standard of ptosis diagnosis), and (3) we also calculate supporting features such as percentage of iris visible through the eyelids. Features such as the latter were too difficult to calculate before, but with machine learning, can be an automatic output from any front-facing photograph. Calculating novel metrics like these can help expand clinical workflow and eliminate the need for complex clinical imaging set-ups, potentially supporting efforts in telemedicine.

Project lead: Pete Setabutr

Team Lead: Darvin Yi

Team: Abdullah Aleem, Manoj Prabhakar Nallabothula

The AI Utils project aims to create basic deep learning AI tools for internal research in UIC’s department of ophthalmology and bioengineering. This is done through an amalgamation of API’s, protocols, and software packages hosted in the AI Utils repository . Through dockering workflows, we will define potential folder hierarchies and easy to load models, based on current state of the art (SOTA) performances. For new projects, users will have to simply format their data according to AI Utils protocol, and training should become as simple as a single line command. Upon training a network, AI Utils will also provide an API for future inference calls through docker as well, which will enable easy deployment on cloud compute clusters or even our own backend servers. By well defining and documenting agreements between users and developers, we can strive to democratize AI.