Auxiliary meta-learning strategy for cancer recognition: leveraging external data and optimized feature mapping

Area of Science:

• Oncology
• Computer Science
• Artificial Intelligence

Background:

• Cancer poses a significant global health challenge, with millions of new cases and fatalities annually.
• Deep learning models are effective for cancer recognition but typically require large labeled datasets.
• Few-shot learning methods aim to reduce data requirements but have shown limitations in cancer recognition tasks.

Purpose of the Study:

• To propose an auxiliary meta-learning strategy to enhance few-shot learning for cancer recognition.
• To address the challenges of data scarcity and improve the performance of deep learning models in this domain.
• To develop a method that achieves superior accuracy and generalization capabilities in cancer recognition.

Main Methods:

• An auxiliary meta-learning strategy involving auxiliary training with external data to neutralize misclassification probabilities.
• Feature mapping model optimization using category prototypes and cosine distance in the embedding space.
• Utilizing depthwise over-parameterized convolutional layers and a three-branch structure for accelerated training and performance enhancement.

Main Results:

• The proposed method demonstrated superior accuracy in cancer recognition across BreakHis, Pap smear, and ISIC 2018 datasets.
• Experiments on few-shot benchmark datasets confirmed the approach's excellent generalization capabilities.
• The strategy effectively reduces the need for extensive labeled samples in cancer recognition tasks.

Conclusions:

• The auxiliary meta-learning strategy significantly improves few-shot cancer recognition performance.
• The method offers a promising solution for overcoming data limitations in medical image analysis.
• The approach exhibits strong generalization, making it applicable to various cancer datasets.