HistoEM: A Pathologist-Guided and Explainable Workflow Using Histogram Embedding for Gland Classification.

Area of Science:

• Digital Pathology
• Computational Biology
• Oncology

Background:

• Pathologists utilize established criteria for prostate cancer diagnosis and grading.
• Current convolutional neural networks (CNNs) for prostate cancer detection do not integrate this pathologist-derived knowledge.
• The alignment between machine learning-derived features and pathologist diagnostic features remains unclear.

Purpose of the Study:

• To develop a framework that trains algorithms to discern cellular and subcellular distinctions between benign and cancerous prostate glands.
• To investigate if machine learning models learn features consistent with those used by pathologists in prostate cancer grading.

Main Methods:

• A novel framework, HistoEM, was developed for analyzing hematoxylin and eosin-stained digital prostate tissue slides.
• The pipeline involves accurate gland segmentation, stroma exclusion, and utilizes histogram embedding of CNN latent space features.
• A two-stage network classifies glands as benign vs. cancer and further grades cancerous glands (low vs. high grade) using U-Net architecture.

Main Results:

• The HistoEM model achieved performance comparable to state-of-the-art prostate cancer grading models at the gland level.
• Feature analysis revealed that HistoEM prioritizes nuclear features, aligning with pathologist diagnostic practices.
• Visualization techniques like Grad-CAM confirmed the model's focus on relevant nuclear characteristics.

Conclusions:

• The HistoEM framework successfully integrates pathologist-recognized features into CNNs for prostate cancer detection and grading.
• The model's reliance on nuclear features demonstrates a convergence with human expert interpretation.
• This approach offers a broadly applicable method for visualizing and understanding computer-learned features in histopathology.