Representative Data Selection for Efficient Medical Incremental Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural network training requires substantial annotated data, which is often challenging to acquire at once in fields like medical imaging and industrial defect detection.
• New data frequently arrives incrementally, necessitating models that can adapt and improve over time without complete retraining.

Purpose of the Study:

• To develop a data selection strategy for efficient incremental learning, enabling models to retrain with a fixed amount of data as new information becomes available.
• To enable continuous model improvement and adaptation to new data while preserving previously learned information.

Main Methods:

• A hybrid approach combining a variational autoencoder (VAE) with an adversarial network for intelligent data selection.
• Constraining retraining to a fixed subset of data to achieve fast model updates.
• Validation on the LGG Segmentation dataset for a semantic segmentation task.

Main Results:

• The VAE-based data selection model combined with adversarial training effectively selects representative and reliable data subsets.
• Achieved time-efficient incremental learning, allowing for rapid model retraining and adaptation.
• Demonstrated the model's ability to continually learn from small, incoming data batches without catastrophic forgetting.

Conclusions:

• The proposed framework offers a practical solution for incremental learning in data-scarce scenarios, particularly in medical image analysis.
• Enables immediate visualization of model improvements as new annotated data becomes available, facilitating quicker clinical relevance.
• Facilitates efficient adaptation of deep learning models to evolving datasets through smart data selection and retraining.