MedOptNet: Meta-Learning Framework for Few-Shot Medical Image Classification.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Science

Background:

• Limited data and high annotation costs pose challenges for medical image classification.
• Few-shot learning is crucial for developing AI models in data-scarce medical research.

Purpose of the Study:

• To propose MedOptNet, a meta-learning framework for efficient few-shot medical image classification.
• To integrate high-performance convex optimization models as classifiers within the framework.

Main Methods:

• Developed a meta-learning framework (MedOptNet) for few-shot medical image classification.
• Utilized convex optimization models (e.g., kernel SVM, ridge regression) as classifiers.
• Implemented end-to-end training via dual problems and differentiation, incorporating regularization techniques.

Main Results:

• MedOptNet demonstrated superior performance compared to benchmark models on BreakHis, ISIC2018, and Pap smear datasets.
• The framework achieved effective classification with limited medical data.
• Ablation studies confirmed the effectiveness of individual modules within MedOptNet.

Conclusions:

• MedOptNet offers an effective solution for few-shot medical image classification challenges.
• The framework's ability to leverage convex optimization models enhances its performance and generalization.
• MedOptNet shows promise for advancing AI applications in medical research with limited data.