Temel Kavramlar
SAMDA combines SAM with nnUNet for efficient domain adaptation in EM segmentation.
Özet
Traditional deep learning methods struggle with limited samples and annotations in EM segmentation.
Large-scale vision models face challenges in fine-tuning for domain transfer.
SAMDA integrates SAM and nnUNet to enhance transferability and accuracy.
The model improves mitochondria segmentation by 6.7% with only a single annotated image.
Tested on electron microscopic and MRI datasets, demonstrating generalization ability.
Introduction:
Mitochondria segmentation is crucial for cellular understanding.
Challenges include variable size, shape, and distribution of mitochondria.
Methodology:
UNet model enables dense fusion of shallow and deep features.
SAM-based adaptation module enhances robustness and transferability.
Experimental setup:
Experiments conducted on EM datasets (Kasthuri++, EPFL) and MRI datasets (HarP, Hammers, Oasis, LPBA40).
Models implemented in PyTorch with specific training settings.
Result and Discussion:
SAM-based encoders outperform nnUNet under various few-shot conditions.
MedSAM encoder shows the best performance in domain adaptation tasks.
Conclusion:
SAMDA effectively integrates SAM with nnUNet for high transferability in EM segmentation.
Demonstrated notable performance improvements across different datasets.
İstatistikler
"The effectiveness of our model is evaluated on two electron microscopic image datasets with different modalities for mitochondria segmentation, which improves the dice coefficient on the target domain by 6.7%."
"Our model is further verified on four MRI datasets from different sources to prove its generalization ability."