Meta-learning with Unlabeled Query Updating and Consistency Learning for Few-shot OCT Image Classification

Objective: Deep neural networks are widely used in the field of optical coherence tomography (OCT) to screen some common retinal diseases. However, for rare diseases with fewer cases for model training, it is challenging to achieve automatic diagnosis using traditional deep learning. Meta-learning based few-shot learning can be used to address the problem of insufficient training data. Methods: We propose a novel algorithm for few-shot OCT image classification, where meta-learning is used to fine-tune the pre-trained model and obtain good initialization for task generalization. Unsupervised learning based on query data is for the first time introduced in meta-learning. Cross-set consistency learning is proposed to reduce the gap between meta-knowledge learned from support and query data. Data mixup is also integrated to generate virtual samples and enhance data variety. Results: A lightweight subset was constructed based on a public OCT dataset and extensive experiments were performed. The classification accuracy of the proposed method was higher than existing few-shot learning methods. To show the generalization of the proposed method, experiments were also performed on a histological image dataset, and superior performance was also achieved. Conclusion: The proposed strategies help the model to fully utilize the limited data and to explore hidden information, improving its generalization to unseen tasks. Significance: The proposed method has great value in training deep learning models for diagnosis of rare diseases.