Meta-Learning With Unlabeled Query Updating and Consistency Learning for Few-Shot OCT Image Classification

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are vital for diagnosing common retinal diseases using optical coherence tomography (OCT).
• Diagnosing rare retinal diseases with DNNs is challenging due to insufficient training data.
• Meta-learning based few-shot learning offers a solution for data-scarce scenarios.

Purpose of the Study:

• To develop a novel algorithm for few-shot OCT image classification.
• To address the challenge of diagnosing rare diseases with limited OCT data.
• To improve the generalization capabilities of deep learning models for rare disease diagnosis.

Main Methods:

• A meta-learning algorithm fine-tunes pre-trained models for task generalization.
• Unsupervised learning on query data is integrated into meta-learning.
• Cross-set consistency learning minimizes discrepancies between support and query data.
• Data mixup generates virtual samples to increase data diversity.

Main Results:

• The proposed method achieved higher classification accuracy than existing few-shot learning techniques on an OCT dataset.
• Experiments on a histological image dataset demonstrated superior performance, confirming generalization.
• The algorithm effectively utilizes limited data and uncovers hidden information.

Conclusions:

• The developed strategies enhance model performance by maximizing the utility of limited data.
• The novel approach shows significant value for training deep learning models in rare disease diagnosis.
• This method improves model generalization to previously unseen tasks.