Toward robust histopathology imaging: An unsupervised framework for artifact detection, localization, and restoration

Area of Science:

• Digital pathology
• Computational histopathology
• Medical image analysis

Background:

• Whole Slide Images (WSIs) are crucial for modern pathology.
• Image artifacts from tissue processing and scanning degrade WSI quality.
• Existing artifact correction methods often require annotated data and lack precise localization.

Purpose of the Study:

• To develop an unsupervised pipeline for automated artifact detection, localization, and restoration in histopathology images.
• To improve the robustness and efficiency of computer-aided diagnosis (CAD) systems.
• To provide an automated solution for artifact management in digital pathology workflows.

Main Methods:

• Feature extraction using a histopathology foundation model.
• Anomaly heatmap generation via Normalizing Flow for artifact detection.
• Unsupervised semantic segmentation for artifact localization.
• Mask-guided artifact restoration using a diffusion model.

Main Results:

• Effective handling of synthetic and real-world artifacts using only normal training data.
• Demonstrated improvement in downstream tasks: BCSS segmentation and TCGA-BLCA tumor staging.
• Successful artifact detection, localization, and restoration without manual annotation.

Conclusions:

• The proposed unsupervised pipeline offers an effective solution for artifact management in histopathology images.
• The method enhances the performance and reliability of computer-aided diagnosis (CAD) systems.
• This approach reduces the need for manual verification and streamlines digital pathology workflows.