Decomposition based curriculum-style self-training for source-free universal domain adaptation in computational...

Area of Science:

• Digital pathology
• Machine learning in medicine
• Computer-aided diagnosis

Background:

• Computational pathology models aid tissue typing and whole slide image analysis.
• Cross-institute deployment is hindered by data regulations and domain shifts.
• Source-free universal domain adaptation (SF-UniDA) enables knowledge transfer with unlabeled target data.

Purpose of the Study:

• To address limitations in existing SF-UniDA methods that overlook internal data variations and noise.
• To propose a novel framework for improved domain adaptation in computational pathology.
• To enhance the cross-institute deployment of pathology models.

Main Methods:

• Developed a Decomposition based Curriculum-style Self-Training (DCST) framework.
• Introduced an adaptive division strategy using Gaussian Mixture Models for easy/hard data separation.
• Implemented prototype-aware regularization and alignment for maintaining inter-class variability and balancing learning.

Main Results:

• The DCST framework effectively handles internal variations within private data.
• The proposed method mitigates interference from noisy samples.
• Consistent performance gains were achieved on colorectal cancer phenotyping datasets compared to state-of-the-art methods.

Conclusions:

• DCST offers a robust approach to SF-UniDA by improving data scrutiny and learning paradigms.
• The framework enhances the reliability and applicability of computational pathology models across different institutions.
• This work advances the field of AI in digital pathology for clinical applications.