Specialized AI Outperforms GPT-5 and Gemini in Medical Imaging: Mayo Clinic's Domain-Focused Approach Challenges Big Tech's Generalist Models
LENS, a fine-tuned endoscopic AI from Mayo Clinic, achieved 92.8% accuracy while operating faster and cheaper than leading foundation models, suggesting specialized medical AI may trump general-purpose systems.

The race to dominate medical AI may not be won by the biggest models. Researchers at Mayo Clinic have developed LENS (Large language model-powered Endoscopic Narrative System), a specialized vision-language AI that outperformed 22 leading generalist models—including GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5—in endoscopic interpretation tasks, while operating at significantly lower cost and latency.
The findings challenge the prevailing assumption that frontier foundation models will dominate clinical applications, instead suggesting that domain-specific fine-tuning may offer a more practical pathway for medical AI deployment.
A Single Model for Multiple Clinical Tasks
LENS achieved 92.8% accuracy and a 0.805 macro-F1 score across 10 clinically relevant endoscopic interpretation tasks. The model was developed by a team led by Yichen Wang, Yuting Huang, Cui Tao, Vivek Kumbhari, and Yutao Zou at Mayo Clinic Florida's GI and Artificial Intelligence and Informatics division, under the direction of Michael Wallace, Fred C. Andersen Endowed Professor of Medicine.
What distinguishes LENS from previous approaches is its multitask architecture. Rather than deploying separate convolutional neural networks (CNNs) for each specific diagnostic task—the traditional approach in medical imaging AI—LENS handles diverse interpretation challenges through a single fine-tuned model. This design not only matched or exceeded the performance of task-specific CNNs but did so while generating narrative reports alongside lesion recognition, potentially standardizing documentation and reducing physician reporting burden.
The model was fine-tuned on curated endoscopic images as a domain-specific vision-language system, optimizing it specifically for gastrointestinal imaging rather than attempting to serve all medical imaging modalities or general visual tasks.
Efficiency Advantages Over Foundation Models
Beyond raw accuracy, LENS demonstrated substantial practical advantages. The specialized model achieved its superior performance at a fraction of the inference latency and cost compared to generalist models like GPT-5 and Gemini 2.5 Pro. This efficiency gap has significant implications for real-world deployment, where computational costs and response times directly impact clinical workflow and economic viability.
The research team acknowledges an important limitation: all evaluations were performed on curated still images, and they anticipate a performance drop in real-world clinical practice where image quality, patient movement, and procedural variability introduce additional challenges. Nevertheless, the benchmark results suggest that specialized, domain-tuned models can compete effectively without requiring the massive infrastructure of frontier foundation models.
Strategic Implications for Medical AI Development
Mayo Clinic's findings support a strategic direction favoring fine-tuned generalist models over task-by-task CNN development as the future pathway for AI in endoscopy. This approach offers healthcare institutions and medical AI companies a middle ground: models that are specialized enough to achieve clinical-grade performance in specific domains, yet flexible enough to handle multiple related tasks without requiring separate systems for each application.
For health tech companies, this validates the commercial viability of domain-specific models that balance performance with computational efficiency. For healthcare providers, LENS represents a potential paradigm shift—a single multitask model that could streamline workflow and reduce the need to deploy multiple specialized AI tools, though real-world validation studies remain necessary to confirm clinical utility and patient outcomes.
Sources
Follow Health AI Central
Keep up with health AI where you already read
On LinkedIn
The day’s strongest health AI stories in your feed, one or two each weekday.
On Google
Choose Health AI Central as a preferred source, and Google is more likely to show our reporting in Top Stories.
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.