Incremental Learning for Heterogeneous Structure Segmentation in Brain Tumor MRI.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computer vision

Background:

• Static deep learning (DL) models excel in single-domain segmentation but falter in evolving environments.
• Incremental learning faces challenges like distribution shifts and new structures without source data.
• Catastrophic forgetting hinders model updates in dynamic medical data scenarios.

Purpose of the Study:

• To develop a unified framework for progressively evolving off-the-shelf segmentation models to diverse datasets with new anatomical categories.
• To enable lifelong learning for segmentation models in the context of accumulating big medical data.
• To address challenges of distribution shifts and incremental structures in continual learning settings.

Main Methods:

• Proposed a divergence-aware dual-flow module with rigidity and plasticity branches, guided by continuous batch renormalization.
• Developed a complementary pseudo-label training scheme with self-entropy regularized momentum MixUp decay for adaptive optimization.
• Evaluated the framework on brain tumor segmentation with changing MRI scanners/modalities and incremental structures.

Main Results:

• The framework successfully evolved segmentation models to diverse datasets with additional anatomical categories.
• Demonstrated retention of discriminability for previously learned structures, mitigating catastrophic forgetting.
• Achieved effective lifelong model extension in a dynamic brain tumor segmentation task.

Conclusions:

• The proposed framework offers a unified approach for adapting static DL segmentation models to evolving medical data.
• It enables realistic lifelong extension of segmentation models, crucial for handling large-scale, continuously accumulating medical data.
• The method effectively balances model plasticity and rigidity for robust incremental learning in medical imaging.