Few-Shot Learning for Medical Image Segmentation Using 3D U-Net and Model-Agnostic Meta-Learning (MAML)

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning for image segmentation requires extensive annotated data, which is scarce in medical imaging.
• Few-shot learning techniques offer a solution by leveraging prior knowledge to generalize from limited samples.

Purpose of the Study:

• To adapt Model-Agnostic Meta-Learning (MAML), a gradient-based meta-learning algorithm, for efficient medical image segmentation.
• To evaluate the performance of MAML combined with an enhanced 3D U-Net on segmenting multiple abdominal organs using limited annotated data.

Main Methods:

• Utilized Model-Agnostic Meta-Learning (MAML) for rapid adaptation to new segmentation tasks.
• Employed an enhanced 3D U-Net as the base convolutional neural network architecture.
• Evaluated the approach on the TotalSegmentator dataset and a local hospital dataset across various few-shot settings (5-shot and 10-shot).

Main Results:

• Achieved high mean Dice coefficients in few-shot settings on the TotalSegmentator dataset: up to 93.70% (liver, 10-shot) and 90.27% (liver, 5-shot).
• Demonstrated strong performance on a local hospital dataset, with 5-shot segmentation achieving up to 90.62% (liver).
• The MAML-based approach showed effective rapid adaptation for segmenting liver, spleen, and kidneys with minimal annotated images.

Conclusions:

• Model-Agnostic Meta-Learning (MAML) combined with an enhanced 3D U-Net is effective for medical image segmentation with limited data.
• The proposed method facilitates rapid generalization to new segmentation tasks, addressing data scarcity challenges in clinical settings.
• This approach holds significant potential for improving the efficiency and applicability of AI in medical image analysis.