Accurate segmentation of prostate cancer histomorphometric features using a weakly supervised convolutional neural...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in histopathology

Background:

• Prostate cancer originates from glandular epithelium, necessitating accurate histological assessment.
• Traditional morphometric techniques for analyzing glandular epithelium in automated pipelines are often rigid and struggle with dataset variations.
• Existing methods face challenges in robustness due to staining, imaging, and preparation inconsistencies in large datasets.

Purpose of the Study:

• To quantify the performance of a pixelwise segmentation algorithm for prostate histology.
• To evaluate an algorithm trained with combined weak and strong labels for stroma, epithelium, and lumen (SEL) regions.
• To improve automated detection and classification of prostate cancer in histological images.

Main Methods:

• Trained a convolutional neural network (CNN) using a combination of weakly labeled datasets (from morphometrics) and high-quality labeled datasets (from human observers).
• Applied the trained CNN for pixelwise segmentation of stromal, epithelium, and lumen (SEL) regions in prostate biopsy cores and whole-mount tissues.
• Utilized H&E-stained tissue samples for analysis.

Main Results:

• Training a deep learning algorithm on weakly labeled data enhanced classification robustness compared to traditional morphometric methods.
• The developed algorithm achieved improved qualitative SEL labeling in prostate tissue.
• Deep learning-generated labels demonstrated superior performance for cancer classification in higher-order algorithms compared to morphometrically derived labels.

Conclusions:

• Combining weak and strong labels effectively trains CNNs for improved SEL labeling in prostate histology.
• The deep learning-based segmentation approach offers superior performance for cancer classification over traditional methods.
• This method enhances the accuracy and robustness of automated analysis in digital pathology pipelines.