Privacy-preserving continual learning methods for medical image classification: a comparative analysis.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• Deep learning models for medical image classification face challenges like performance degradation and limited adaptability.
• Frequent retraining is unfeasible and raises privacy concerns due to patient data retention.

Purpose of the Study:

• To investigate privacy-preserving continual learning methods as an alternative to frequent retraining.
• To evaluate the effectiveness of these methods in medical image classification tasks.

Main Methods:

• Evaluated twelve privacy-preserving non-storage continual learning algorithms using deep learning models.
• Classified retinal diseases from optical coherence tomography (OCT) images in a class-incremental learning scenario.
• Tested algorithms on colon cancer histology and CIFAR10 datasets for proof of concept and benchmark comparison.

Main Results:

• Brain-inspired replay (BIR) achieved the highest accuracy (62.00%) for retinal disease classification from OCT images.
• Efficient Feature Transformations (EFT) attained the highest accuracy (66.82%) for colon cancer histology classification.
• The finetune model without continual learning exhibited catastrophic forgetting, while joint retraining models showed superior performance.

Conclusions:

• Continual learning methods show promise in mitigating catastrophic forgetting and enabling continuous model updates.
• These methods are crucial for preserving privacy in healthcare deep learning models.
• Privacy-preserving continual learning is a promising solution for long-term clinical deployment of AI models.