Ilya Pershin
Head, Med AI Lab · Innopolis University, Russia · PhD
My research centres on human attention as a measurable, usable signal in clinical work. A clinician reading an image is already producing data about where the diagnostic information is — and about their own state while producing it. That signal can be used in both directions: to assess the quality of the reading, and to supervise models more cheaply than manual annotation allows. Main lines of interest:
- Expert gaze and human–AI interaction: eye-tracking experiments with practising radiologists; gaze patterns linked to fatigue, reading quality and expertise; how AI changes diagnostic quality, and what effective clinician–AI interaction looks like
- Gaze-based segmentation: a clinician’s eye movements as prompts and as a correction mechanism for interactive medical image segmentation, including volumetric data
- Gaze modelling: generating synthetic gaze data for reading text and viewing images; using synthetic gaze to give neural networks information about human attention
PhD in Engineering Sciences (2025, speciality 1.2.1 — Artificial Intelligence and Machine Learning); dissertation: Deep Models for Medical Diagnosis Based on Radiologist Attention.
The lab also works on electronic health records and IVF embryo assessment — see the lab page.
Prospective students
In 2025–2026 the lab produced 13 papers at A* venues, two of them in main track (EMNLP, NeurIPS), and two Q1 journal papers; several undergraduates are co-authors. Recent lab members have gone on to the University of Copenhagen and the University of Amsterdam. How to join is on the lab page.
News
- 2026
- Multilingual Synthetic Scanpaths — ICML 2026 SPIGM workshop, Seoul.
- 2026
- Cross-Lingual Transfer Learning for Enhanced Synthetic Scanpath Prediction — AAAI 2026 AIBSD workshop, Singapore.
- 2026
- Simulating Concept Bottlenecks with Vision Language Models — ICLR 2026 UCRL workshop, with Karim Galliamov and Ivan Titov.
- Dec 2025
- PhD defended: Deep models for medical diagnosis based on radiologist attention, FRC “Computer Science and Control” RAS, Moscow.
- 2025
- Enhancing RLHF with Human Gaze Modeling — EMNLP 2025, main track.
- 2025
- From Human Attention to Diagnosis — poster at NeurIPS 2025, San Diego.
- 2025
- Two Q1 papers in IEEE Access: gaze-assisted segmentation correction, and uncertainty in cardiac landmark detection.