Incremental Learning Meets Transfer Learning: Application to Multi-site Prostate MRI Segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Medical image segmentation tasks benefit from large datasets.
• Existing multi-site training methods require all data simultaneously, limiting practical deployment.
• Sequential training of a single model for improved performance and generalization is desired.

Purpose of the Study:

• To propose a novel incremental-transfer learning (ITL) framework for sequential multi-site medical image segmentation.
• To enable a single model to perform better across datasets and generalize to new domains.
• To address catastrophic forgetting in incremental learning.

Main Methods:

• Developed an end-to-end sequential training framework (ITL).
• Utilized a site-agnostic encoder with pretrained weights and multiple decoder heads.
• Introduced a site-level incremental loss for improved generalization.
• Leveraged linear combinations of embedding features for knowledge transfer.

Main Results:

• Demonstrated ITL's effectiveness in alleviating catastrophic forgetting.
• Validated the approach on five challenging benchmark datasets.
• Achieved improved performance and generalization compared to existing methods.

Conclusions:

• ITL provides a robust solution for sequential multi-site medical image segmentation.
• The framework minimizes assumptions on computational resources and expertise.
• ITL offers a strong foundation for future advancements in the field.