First-Stage Prostate Cancer Identification on Histopathological Images: Hand-Driven versus Automatic Learning

Area of Science:

• Digital pathology
• Computational oncology
• Machine learning in medicine

Background:

• Accurate prostate cancer identification relies on histopathological image analysis.
• Current research often focuses on Gleason grading, but early-stage discrimination is crucial.
• Automated gland segmentation provides a basis for analyzing tissue characteristics.

Purpose of the Study:

• To develop a method for discriminating between healthy and early-stage cancerous prostate tissues using segmented gland candidates.
• To compare a novel hand-driven feature extraction approach with deep learning methods for this discrimination task.

Main Methods:

• Extracted hand-crafted features including morphology, texture, fractals, and contextual information from segmented gland candidates.
• Performed statistical analysis to select the most relevant features for machine learning classifiers.
• Applied deep learning, specifically fine-tuning the VGG19 neural network's convolutional block, to gland images.
• Utilized a nonlinear Support Vector Machine (SVM) for the hand-driven learning approach.

Main Results:

• The hand-driven learning approach with a nonlinear SVM achieved the highest multi-class accuracy of 0.876 ± 0.026.
• This accuracy was obtained in discriminating between artefacts, benign glands, and Gleason grade 3 cancerous glands.
• The study demonstrated the effectiveness of combining diverse feature descriptors and statistical selection.

Conclusions:

• The proposed hand-driven learning approach, integrating comprehensive feature extraction and optimized machine learning, shows significant promise for early prostate cancer detection.
• Deep learning methods, while explored, did not outperform the optimized hand-driven approach in this specific task.
• This work highlights the potential of analyzing segmented gland candidates for accurate tissue classification in digital pathology.