Cervical cytopathology image refocusing via multi-scale attention features and domain normalization

Area of Science:

• Digital Pathology
• Medical Image Processing
• Computational Cytology

Background:

• Defocus blur in cervical cytopathology whole slide images (WSIs) hinders accurate analysis.
• Existing deblurring methods often target global motion blur and require extensive retraining for new datasets.
• Local defocus blur in cytopathology images presents unique challenges not adequately addressed by current techniques.

Purpose of the Study:

• To develop an effective refocusing method for cervical cytopathology images, specifically targeting local defocus blur.
• To create a method robust to unseen domains without requiring extensive supervised retraining.
• To demonstrate the improvement in image quality and subsequent analysis task performance using the proposed method.

Main Methods:

• Proposed a two-part method: a domain normalization net (DNN) and a refocusing net (RFN).
• DNN utilizes a registration-free cycle scheme with gray mask loss and hue-encoding mask loss for domain normalization.
• RFN employs a multi-scale refocusing network with a defocus intensity estimation mask to reconstruct cell structures and local blur.

Main Results:

• The proposed method achieves superior refocusing performance on cervical cytopathology images compared to state-of-the-art deblurring models.
• Hybrid learning strategy enables effective refocusing on unsupervised domains.
• Refocused images significantly improve the performance of subsequent high-level analysis tasks.

Conclusions:

• The developed multi-scale attention and domain normalization method effectively addresses local defocus blur in cervical cytopathology images.
• The method demonstrates robustness across different domains, reducing the need for extensive retraining.
• The released dataset and source code will facilitate further research in cytopathology image refocusing.