TinyMIG: Transferring Generalization from Vision Foundation Models to Single-Domain Medical Imaging

Chuang LIU, Hongyan Xu, Yichao Cao, Xiu Su, Zhe Qu, Tianfa Li, Shan An, Haogang Zhu

Abstract: Medical imaging faces significant challenges in single-domain generalization (SDG) due to the diversity of imaging devices and the variability among data collection centers. To address these challenges, we propose \textbf{TinyMIG}, a framework designed to transfer generalization capabilities from vision foundation models to medical imaging SDG. TinyMIG aims to enable lightweight specialized models to mimic the strong generalization capabilities of foundation models in terms of both global feature distribution and local fine-grained details during training. Specifically, for global feature distribution, we propose a Global Distribution Consistency Learning strategy that mimics the prior distributions of the foundation model layer by layer. For local fine-grained details, we further design a Localized Representation Alignment method, which promotes semantic alignment and generalization distillation between the specialized model and the foundation model. These mechanisms collectively enable the specialized model to achieve robust performance in diverse medical imaging scenarios. Extensive experiments on large-scale benchmarks demonstrate that TinyMIG, with extremely low computational cost, significantly outperforms state-of-the-art models, showcasing its superior SDG capabilities. All the code and model weights will be publicly available.

Lay Summary:

Medical images, such as scans from hospitals, often look different depending on the machines used and where they were taken. This makes it hard for AI tools trained on one dataset to work well on new, unseen data — a problem known as single-domain generalization.

We introduce \textbf{TinyMIG}, a method that helps small, efficient AI models learn from powerful, general-purpose vision models (like those behind tools such as ChatGPT or DALL·E). Our method teaches the smaller model to copy how the big model understands both the overall structure and fine details of images during training.

To do this, TinyMIG aligns the way the small model “sees” image features with how the big model does, both at a broad and detailed level. This helps the small model adapt better to new types of medical images.

Our results show that TinyMIG can outperform other approaches — all while using much less computing power — making it practical for real-world healthcare use.

Primary Area: Applications->Health / Medicine

Keywords: medical image segmentation, generalization, vision foundation models

Submission Number: 1041