Computer-Assisted Annotation of Digital H&E/SOX10 Dual Stains Generates High-Performing Convolutional Neural Network...

Area of Science:

• Digital pathology
• Computational oncology
• Machine learning in histopathology

Background:

• Deep learning for H&E stain analysis necessitates extensive annotated datasets.
• Pathologist annotation is time-consuming and requires specialized expertise.

Purpose of the Study:

• To optimize and evaluate computer-assisted annotation using digital dual stains.
• To develop a convolutional neural network (CNN) for accurate tumor burden calculation in melanoma.

Main Methods:

• Digitized H&E stains of melanoma, re-stained with SOX10, and re-scanned.
• Aligned images allowed direct transfer of SOX10 annotations to H&E stains.
• Developed a CNN for tumor burden (CNN_TB) based on over 1.2 million annotated nuclei.

Main Results:

• High annotation precision for tumor and normal cells in primary melanomas.
• Annotation precision for normal cells was reduced in metastases due to SOX10 variability.
• CNN_TB showed a smaller mean difference in tumor burden compared to pathologists for skin lesions.

Conclusions:

• Computer-assisted annotation with dual stains efficiently creates large, high-quality H&E training sets for melanoma.
• The developed CNN_TB is a high-performing tool for tumor burden assessment, outperforming pathologists for primary and subcutaneous metastases.