MULTI-DOMAIN LEARNING BY META-LEARNING: TAKING OPTIMAL STEPS IN MULTI-DOMAIN LOSS LANDSCAPES BY INNER-LOOP LEARNING.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Existing Multi-Domain Learning (MDL) techniques often require model-specific architectural changes, complicating integration with established models like U-Net.
• Applying MDL to multi-modal data, such as neuroimaging, necessitates efficient and adaptable solutions.
• The need for model-independent methods is crucial for leveraging diverse datasets without extensive re-engineering.

Purpose of the Study:

• To develop a model-agnostic algorithmic solution for Multi-Domain Learning (MDL) in multi-modal applications.
• To enable widely used neural networks to perform MDL without requiring architectural modifications.
• To apply the proposed method to the automatic segmentation of white matter hyperintensity (WMH) in medical imaging.

Main Methods:

• A weighted loss function is extended using meta-learning techniques, specifically inner-loop gradient steps.
• The method dynamically estimates posterior distributions over loss function hyperparameters.
• The approach is purely algorithmic, requiring no additional model parameters or network architecture changes.

Main Results:

• The proposed model-agnostic method successfully enables Multi-Domain Learning (MDL) through algorithmic modifications.
• Demonstrated effectiveness in the medical imaging application of automatic white matter hyperintensity (WMH) segmentation.
• Utilized complementary information from two neuroimaging modalities (T1-MR and FLAIR) for improved performance.

Conclusions:

• The developed algorithmic approach provides a versatile and efficient solution for Multi-Domain Learning (MDL).
• This model-agnostic strategy simplifies the application of MDL to various neural networks and multi-modal datasets.
• The method shows significant promise for medical image analysis tasks, particularly in segmenting WMH.