DiagSet: a dataset for prostate cancer histopathological image classification

Area of Science:

• Oncology
• Computational Pathology
• Machine Learning

Background:

• Prostate cancer detection presents significant challenges in histopathology.
• Accurate and efficient diagnostic tools are crucial for patient outcomes.

Purpose of the Study:

• To introduce a novel, large-scale histopathological dataset for prostate cancer detection.
• To propose and evaluate a machine learning framework for identifying cancerous regions and predicting diagnoses from histopathological scans.

Main Methods:

• Development of a comprehensive dataset with over 2.6 million tissue patches from annotated scans.
• Implementation of an ensemble deep neural network framework for patch-level recognition and scan-level diagnosis.
• Utilizing thresholding to manage uncertain cases and abstaining from decisions when necessary.

Main Results:

• The proposed machine learning framework achieved 94.6% accuracy in patch-level cancer recognition.
• Scan-level diagnosis predictions demonstrated high statistical agreement with human expert histopathologists.
• The dataset is publicly available for further research and development.

Conclusions:

• The novel dataset and machine learning framework show significant promise for improving prostate cancer detection accuracy.
• The developed approach offers a robust tool for computational pathology, potentially assisting in clinical decision-making.
• Further validation and application of this framework can enhance diagnostic efficiency and reliability in prostate cancer diagnosis.