Interactive Few-Shot Learning: Limited Supervision, Better Medical Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised deep learning for medical image segmentation requires extensive data annotation.
• Few-shot segmentation methods reduce annotation burden but often lack adaptability to new tasks.

Purpose of the Study:

• To develop a novel few-shot segmentation approach, Interactive Few-shot Learning (IFSL), that addresses annotation burden and improves task adaptability.
• To introduce a new network architecture (MPrNet) and an interactive test-time optimization algorithm (IL-TTOA).

Main Methods:

• Designed Medical Prior-based Few-shot Learning Network (MPrNet) using minimal annotated samples (e.g., 10) for guidance without pre-training.
• Proposed Interactive Learning-based Test Time Optimization Algorithm (IL-TTOA) for on-the-fly interactive model strengthening.
• IFSL enables interactive and controllable optimization of few-shot segmentation models on target tasks.

Main Results:

• IFSL approach outperformed state-of-the-art methods by over 20% in Dice Similarity Coefficient (DSC) on four tasks.
• The IL-TTOA algorithm provided an additional ~10% DSC improvement.
• Demonstrated the first interactive and controllable optimization for few-shot segmentation models.

Conclusions:

• IFSL effectively reduces annotation burden in medical image segmentation.
• The proposed method significantly enhances model performance and adaptability for few-shot segmentation tasks.
• Interactive optimization is a promising strategy for improving few-shot learning in medical imaging.