Towards Foundation Models and Few-Shot Parameter-Efficient Fine-Tuning for Volumetric Organ Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Foundation models and the pre-train-and-adapt paradigm are increasingly popular for volumetric medical image segmentation.
• Full fine-tuning strategies for transfer learning can be resource-intensive and perform poorly with limited labeled data, hindering clinical application.

Purpose of the Study:

• To formalize Few-Shot Efficient Fine-Tuning (FSEFT) as a realistic scenario for adapting medical image segmentation foundation models.
• To address the challenges of data and parameter efficiency in adapting foundation models for medical image segmentation.

Main Methods:

• Leveraged a foundation model pre-trained on CT organ segmentation data.
• Employed Parameter-Efficient Fine-Tuning and black-box Adapters for efficient adaptation.
• Introduced novel Spatial black-box Adapters for dense prediction tasks and transductive inference, incorporating task-specific prior knowledge.

Main Results:

• Demonstrated the suitability of foundation models for medical image segmentation.
• Highlighted the limitations of popular fine-tuning strategies in few-shot scenarios.
• Showcased the effectiveness of proposed efficient adaptation methodologies.

Conclusions:

• Foundation models are effective for medical image segmentation, but efficient adaptation techniques are crucial for real-world clinical settings with limited data and resources.
• FSEFT, utilizing parameter-efficient fine-tuning and novel adapters, provides a viable approach for adapting these models under data constraints.