Unsupervised domain adaptive tumor region recognition for Ki67 automated assisted quantification

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Ki67 protein is crucial for assessing breast cancer proliferation and metastasis.
• Accurate tumor region identification on Ki67-stained whole-slide images (WSIs) is vital for clinical prognostication.
• Current deep learning methods for Ki67 quantification require extensive manual annotations, which are time-consuming and labor-intensive.

Purpose of the Study:

• To develop a novel tumor region recognition approach for precise Ki67 quantification.
• To overcome the limitations of massive annotation requirements in deep learning for Ki67 analysis.
• To improve the accuracy and efficiency of Ki67-based prognostic assessments in breast cancer.

Main Methods:

• An unsupervised domain adaptation (UDA) method combining adversarial learning and self-training was proposed.
• The model was trained on labeled hematoxylin and eosin (H&E) data and unlabeled Ki67 data to recognize tumor regions in Ki67 WSIs.
• An automated Ki67 quantification system was developed, incorporating foreground segmentation, tumor region recognition, cell counting, and WSI-level score calculation.

Main Results:

• The UDA method achieved high performance in tumor region recognition, with AUCs of 0.9915 (validation), 0.9352 (internal test), and 0.9689 (external test).
• Performance significantly surpassed baseline methods and rivaled fully supervised approaches.
• Automated quantification on 148 WSIs demonstrated statistical agreement with pathological reports.

Conclusions:

• The proposed UDA method accurately recognizes Ki67 tumor regions, enabling precise quantification.
• The UDA approach is adaptable to other immunohistochemical staining images.
• Automated quantification results are accurate and interpretable, assisting pathologists in diagnosis and interpretation.