To propose a federated learning (FL)-based framework for myopia classification using fundus imaging that enhances cross-center generalization while ensuring data privacy.
Approach:
Framework Design: Developed a privacy-preserving multicenter myopia classification framework allowing joint training of deep learning models without raw data sharing.
Dynamic Aggregation Strategy: Implemented a dynamic aggregation strategy based on federated averaging (FedAvg) to address multicenter data heterogeneity.
Handling Category Imbalance: Incorporated a category imbalance handling mechanism to improve model performance on underrepresented pathological myopia samples.
Model Optimization: Optimized model architecture to adapt to heterogeneous data distributions across different medical centers.
Experimental Validation: Conducted experiments on datasets from multiple medical centers to validate the effectiveness of the proposed methods.
Key Findings:
The proposed FL framework enhances cross-center generalization ability, demonstrated through improved model performance across diverse datasets.
The dynamic aggregation strategy significantly improves model fitness across different data sources, leading to more accurate classifications.
The framework effectively addresses training bias from insufficient samples of pathological myopia, resulting in better recognition of minority categories.
Interpretation:
The research presents a novel solution for diagnosing myopia in a privacy-preserving manner using federated learning.
Limitations:
Challenges remain in achieving consistent image quality across different medical centers, which may affect model accuracy.
Variability in annotation standards across institutions could hinder the model's generalization capabilities.
The scarcity of pathological myopia samples can introduce class imbalance, impacting the model's performance on these categories.
Conclusion:
The study provides technical references for the application of federated learning in medical image analysis, particularly for myopia classification.
