Related Experiment Video
Updated: Jul 24, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
From generalization to precision: A large domain-specific pretrained model for specialized medical tasks
Zhongwen Li1, Yangyang Wang2, Lei Wang3
1Ningbo Key Laboratory of Medical Research on Blinding Eye Diseases, Ningbo Eye Institute, Ningbo Eye Hospital, Wenzhou Medical University, Ningbo 315000, China; National Clinical Research Center for Ocular Diseases, Eye Hospital, Wenzhou Medical University, Wenzhou 325027, China.
A new domain-specific AI model, USPEC, trained on retinal images, outperforms general foundation models in medical diagnostics. This specialized approach enhances accuracy and efficiency in precision medicine applications.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Foundation models offer broad generalizability but may lack precision for specialized medical tasks.
- Precision is critical in medicine, especially for diagnostic and therapeutic decisions in precision medicine.
- Ultra-widefield (UWF) fundus imaging presents unique challenges for AI model development.
Purpose of the Study:
- To investigate if a large, domain-specific pretrained model can surpass general foundation models in specialized medical tasks using UWF imaging.
- To develop and evaluate a novel AI model tailored for UWF fundus image analysis.
- To assess the performance, label efficiency, and training efficiency of the domain-specific model compared to general models.
Main Methods:
- Development of USPEC, a large-scale domain-specific model pretrained on 875,947 unlabeled UWF images via self-supervised learning.
- Adaptation of USPEC to downstream diagnostic and few-shot learning tasks using explicit labels.
- Comparative analysis of USPEC against several general-purpose foundation models (RETFound, RETFound-DE, VisionFM, RetiZero, EyeFound).
- Validation of model robustness on independent, multi-country datasets.
Main Results:
- USPEC consistently outperformed all evaluated general foundation models across various UWF-based diagnostic tasks.
- USPEC demonstrated superior performance in few-shot learning scenarios.
- The domain-specific model exhibited enhanced label and training efficiency compared to general models.
- USPEC's robustness was validated on independent datasets, confirming its generalizability across different geographical origins.
Conclusions:
- Large domain-specific pretrained models represent a powerful paradigm for advancing AI performance in specialized medical fields like ophthalmology.
- USPEC offers a more accurate and efficient approach for AI-driven analysis of UWF fundus images.
- The findings support the development of tailored AI solutions for precision medicine, improving diagnostic accuracy and efficiency.