Мультимодальные модели в медицинской диагностике как универсальный инструмент

Вестник Росздравнадзора. 2026. № 1. С. 14–29.

Назаренко А. Г., Федоров М. В., Moshkin A., Арлазаров В. В., Грибова В. В., Репин Д. А., Вошев Д. В., Клевцова О. Ю., Игнатьев С. А., Люткин Д. А., Romanov A.

Multimodal foundation models and medical multimodal large language models are establishing a new class of diagnostic clinical decision support systems capable of operating on heterogeneous data sources, including medical imaging (X-ray, CT, MRI, ultrasound, histopathology), physiological signals (ECG, EEG), clinical text (electronic health records, reports, discharge summaries), laboratory measurements, molecular profiling data, and related modalities. This article systematizes model architectures and training strategies that enable transferability across tasks and modalities, and discusses requirements for reliability, clinical validation, and regulatory classification of such models. Universality is interpreted as the ability of a single model or a unified modular framework to address a broad spectrum of tasks (detection, segmentation, triage, summarization, information extraction, and vision–language question answering) while preserving auditability of outputs and strict operational constraints. In particular, the system must not issue a final diagnosis or replace the clinician; instead, it provides well-grounded hypotheses, observations, and decision cues suitable for clinical verification and documentation in compliance with existing regulatory requirements.

Keywords: медицинская диагностика, medical diagnostics, foundation models, Фундаментальные модели, multimodal models, мультимодальные модели, SaMD, MDSW, clinical validation, SaMD, MDSW, клиническая валидация