Research map / Label-efficient and robust learning / Recognition and long-tail

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

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Transfers generalization from vision foundation models (SAM, MedSAM, DINOv2) to a lightweight ResNet-34 U-Net for single-source medical image segmentation via distribution-level and frequency-based feature alignment during training only.

Problem

Single-domain (single-source) generalization for medical image segmentation, where models trained on one source site must generalize to unseen imaging devices and centers, while avoiding the high inference cost of foundation models.

Approach

Uses foundation-model feature statistics to modulate a lightweight specialized model's feature distributions (UFDM with a cross-distribution KL consistency loss) and aligns frequency-filtered (FFT amplitude/phase) intermediate features with the foundation model (LRA), so only the small model is used at inference.

Key results

• On fundus segmentation, TinyMIG reaches 84.97% average DSC, versus 81.10%-88.27% for baselines such as DAPSAM (reported as surpassing DAPSAM by 3.92%; table shows 88.27 average for DAPSAM under a different column).
• On prostate segmentation, TinyMIG achieves 83.09% average DSC and 3.06 ASD, vs. 81.31% DSC for DAPSAM, with 22.52M parameters and 22.64 GFLOPs vs. 98.45M and 58.61 GFLOPs for DAPSAM.
• Ablation on prostate shows removing UFDM, CDCL, FDCE/LRA each reduces performance (full 83.09 DSC vs. 81.76 without co-training).

Measured effects reported in the paper

Surprising findings

• Reported averages in the fundus table are internally inconsistent (e.g., the text says 3.92% gain over DAPSAM while the table lists DAPSAM at 88.27 average on a different column), so the headline gap is unclear.

Limitations

• Foundation models are required during training, adding training-time cost and complexity.
• Evaluated only on fundus OD/OC and prostate MRI segmentation benchmarks.

Experimental setup

Datasets: Fundus optic disc/cup segmentation benchmark (Almazroa et al., 2018; Chen et al., 2023) Prostate MRI segmentation benchmark (Chen et al., 2023)

Benchmarks: Fundus OD/OC segmentation (5 domains) Prostate MRI segmentation (6 domains)

Metrics: DSC ASD FLOPs parameters fps

Baselines: ERM MixStyle CSDG MaxStyle EFDM DSU SLAug TriD MAD VPTTA DAPSAM SAMed

Builds on: AdaIN DAPSAM MixStyle SAMed Rein

Closest papers by meaning

• Perturbating, Tuning, and Collaborating: Harnessing Vision Foundation Models for Single Domain Generalization on Medical Imaging (AAAI 2025 · similarity 0.91)
• PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.87)
• MedSG-Bench: A Benchmark for Medical Image Sequences Grounding (NeurIPS 2025 · similarity 0.86)
• How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical Images (ICLR 2026 · similarity 0.85)
• Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.85)
• The Effect of Intrinsic Dataset Properties on Generalization: Unraveling Learning Differences Between Natural and Medical Images (ICLR 2024 · similarity 0.85)