Meta-learning with implicit gradients in a few-shot setting for medical image segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Traditional supervised deep learning for medical imaging requires large datasets, limiting clinical applicability.
• Training separate models for diverse patient populations or lesion types is impractical due to data demands.

Purpose of the Study:

• To introduce an optimization-based implicit model agnostic meta-learning (iMAML) algorithm for few-shot medical image segmentation.
• To enhance model generalization capabilities on unseen datasets with limited training samples.

Main Methods:

• Exploited an optimization-based implicit model agnostic meta-learning (iMAML) algorithm.
• Applied iMAML in few-shot settings for medical image segmentation tasks.
• Evaluated performance on publicly available skin and polyp datasets.

Main Results:

• The proposed iMAML approach significantly outperformed naive supervised baseline models and recent few-shot segmentation methods.
• iMAML demonstrated improved generalization capabilities compared to classical few-shot learning techniques.
• Achieved a 2%-4% improvement in dice score over MAML (Meta-learning Algorithm) in most experiments.

Conclusions:

• iMAML offers a robust solution for medical image segmentation, particularly in few-shot scenarios with limited labeled data.
• This work is the first to apply iMAML to medical image segmentation, showcasing its effectiveness on diverse lesion datasets.
• The method shows promise for practical clinical deployment by addressing data limitations and improving model generalizability.