CCSI: Continual Class-Specific Impression for data-free class incremental learning.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Traditional deep learning struggles with new disease diagnosis due to offline training requirements.
• Class incremental learning addresses new disease classes but suffers from catastrophic forgetting.
• Existing solutions often require storing previous data, raising privacy and storage concerns in healthcare.

Purpose of the Study:

• To propose a novel data-free class incremental learning framework for medical image classification.
• To address catastrophic forgetting without storing historical patient data.
• To enable accurate diagnosis of newly introduced disease types in real-world clinical settings.

Main Methods:

• Developed a data-free framework utilizing synthetic data generation (Continual Class-Specific Impression - CCSI).
• Acquired CCSI via data inversion over gradients, using mean images and continual normalization statistics.
• Updated networks using synthesized data, new class data, and specialized losses (contrastive, margin, cosine-normalized cross-entropy).

Main Results:

• Achieved state-of-the-art performance on four public MedMNIST datasets and in-house echocardiography data.
• Demonstrated significant improvement in classification accuracy, up to 51% over baseline data-free methods.
• Successfully adapted deep learning models to new disease classes without compromising performance on previously learned ones.

Conclusions:

• The proposed data-free incremental learning framework effectively handles new disease classes in medical imaging.
• CCSI synthesis and tailored loss functions mitigate catastrophic forgetting and improve model generalization.
• This approach offers a practical and privacy-preserving solution for continual learning in clinical diagnostics.