Improving unsupervised stain-to-stain translation using self-supervision and meta-learning

Area of Science:

• Digital pathology
• Computational pathology
• Medical image analysis

Background:

• Digital pathology image analysis requires extensive manual annotation due to image variability.
• Unsupervised domain adaptation, particularly image-to-image translation, offers a solution to reduce manual annotation overhead.
• Histological stain variations present a significant challenge for deep learning models in digital pathology.

Purpose of the Study:

• To enable stain-independent applicability of deep learning segmentation models.
• To address variability in histological stains using unsupervised stain-to-stain translation.
• To improve the effectiveness of unsupervised stain-to-stain translation in kidney histopathology.

Main Methods:

• Utilized CycleGANs for stain-to-stain translation in kidney histopathology images.
• Proposed integrating a prior segmentation network for self-supervised, semantic guidance during translation.
• Incorporated extra channels into the translation output to handle meta-information and underdetermined reconstructions.

Main Results:

• The proposed method with semantic guidance achieved the best performance, with instance-level Dice scores between 78% and 92% for most kidney structures.
• CycleGANs showed limited performance for translating structures like arteries.
• Translation performance was generally lower compared to segmentation performed on the original stain.

Conclusions:

• Current unsupervised stain-to-stain translation technologies are unlikely to produce "generally" applicable simulated stains.
• Further advancements are needed to overcome limitations in translating diverse histological structures and achieve stain-independent applicability.