Boosting few-shot confocal endomicroscopy image recognition with feature-level MixSiam

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Confocal laser endomicroscopy (CLE) is a promising early diagnostic tool for gastrointestinal (GI) diseases.
• A significant challenge in CLE is the lack of large-scale annotated datasets, hindering the learning of discriminative features.
• This limitation impacts the development of accurate AI models for GI disease classification.

Purpose of the Study:

• To address the challenge of limited annotated data in CLE for GI diseases.
• To propose a novel method, feature-level MixSiam, for learning discriminative features from CLE images.
• To enhance the classification of gastrointestinal tumors using self-supervised and few-shot learning techniques.

Main Methods:

• The study proposes a two-stage approach: self-supervised learning (SSL) followed by few-shot learning (FS).
• A feature-level feature mixing strategy is introduced within a Siamese network structure during the SSL stage.
• The pre-trained model from SSL is used as a base learner in the FS stage for rapid generalization.

Main Results:

• The feature-level MixSiam method demonstrated improved performance in linear evaluation, outperforming the baseline by 6% (Top-1).
• The proposed method also showed a 2% improvement over supervised models (Top-1) in classification tasks.
• Significant improvements were observed in few-shot classification tasks compared to previous baseline methods.

Conclusions:

• The feature-level MixSiam method effectively learns discriminative features from CLE images, even with limited labeled data.
• The approach enhances the adaptability of Siamese networks to the high intra-class variance in pCLE datasets.
• This method offers a viable solution for improving GI tumor classification in scenarios with scarce annotated CLE data.