Epithelium segmentation using deep learning in H&E-stained prostate specimens with immunohistochemistry as reference...

Area of Science:

• Digital Pathology
• Computational Biology
• Oncology

Background:

• Gland morphology is crucial for prostate cancer (PCa) grading.
• Automated differentiation of epithelial tissue is essential for PCa detection.
• Current manual methods for tissue segmentation can be subjective.

Purpose of the Study:

• To develop a deep learning method for segmenting epithelial tissue in H&E stained prostatectomy slides.
• To establish immunohistochemistry (IHC) as a precise reference standard for segmentation.
• To create a foundation for automated prostate cancer grading pipelines.

Main Methods:

• A U-Net model was trained to segment epithelial structures using IHC staining as ground truth.
• Color deconvolution was used for preprocessing IHC slides.
• A second U-Net was trained on H&E slides using the IHC-derived segmentation.
• The system was validated on an independent external dataset.

Main Results:

• The deep learning system accurately segmented intact glands and individual tumor epithelial cells.
• IHC provided a more objective ground truth than manual outlining, especially for high-grade PCa.
• The method demonstrated strong generalization capabilities on external data.

Conclusions:

• The proposed deep learning method offers accurate epithelial tissue segmentation in prostatectomy slides.
• This technique leverages IHC for robust ground truth generation.
• The segmentation serves as a critical first step towards fully automated PCa grading.