Cross-Boosted Multi-Target Domain Adaptation for Multi-Modality Histopathology Image Translation and Segmentation

Area of Science:

• Digital pathology
• Computational biology
• Medical imaging analysis

Background:

• Current digital pathology workflows primarily utilize single-modality histopathology images.
• This approach overlooks the complementary diagnostic information present in Haematoxylin & Eosin (H&E) and Immunohistochemically (IHC) stained images.

Purpose of the Study:

• To develop a cross-boosted multi-target domain adaptation pipeline for integrating multi-modality histopathology images.
• To enhance cancer diagnosis by leveraging the combined information from H&E and IHC images.

Main Methods:

• Proposed a pipeline comprising Cross-frequency Style-auxiliary Translation Network (CSTN) for image translation and Dual Cross-boosted Segmentation Network (DCSN) for adaptive segmentation.
• CSTN utilizes a Cross-frequency Feature Transfer Module (CFTM) for realistic image generation.
• DCSN incorporates a dual-branch encoder and Bidirectional Cross-domain Boosting Module (BCBM) for cross-modality information complementation.

Main Results:

• Successfully generated H&E and IHC images from fluorescence microscopy images with realistic color and texture.
• Achieved superior multi-target domain adaptive segmentation performance.
• Established the Multi-modality Thymus Histopathology (MThH) dataset, the largest public benchmark for H&E and IHC images.
• Demonstrated outperformance against state-of-the-art methods on translation and segmentation tasks.

Conclusions:

• The proposed pipeline effectively integrates multi-modality histopathology images for improved cancer diagnosis.
• The developed CSTN and DCSN networks offer significant advancements in histopathology image translation and segmentation.
• The MThH dataset provides a valuable resource for future research in multi-modality histopathology analysis.