Deep Convolutional Neural Networks Enable Discrimination of Heterogeneous Digital Pathology Images.

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Pathological evaluation of tumor tissue is crucial for cancer diagnosis.
• Automated image analysis offers potential to improve diagnostic accuracy and reduce human error.

Purpose of the Study:

• To develop and validate a computational pipeline for classifying histopathology images across various cancer types.
• To assess the utility of deep learning models, including Google's Inceptions and ResNet, for cancer subtyping and biomarker identification.

Main Methods:

• Utilized a stand-alone pipeline incorporating convolutional neural networks (CNNs), including basic CNN, Google's Inceptions (V1, V3), and ResNet.
• Employed three training strategies: last layer training, training from scratch, and fine-tuning pre-trained models.
• Applied the pipeline to discriminate lung cancer subtypes, bladder and breast cancer biomarkers, and immunohistochemistry (IHC) staining scores.

Main Results:

• Achieved high accuracies: 100% for cancer tissues, 92% for subtypes, 95% for biomarkers, and 69% for IHC scores.
• Demonstrated the effectiveness of deep learning for identifying cancer subtypes and robustness against tumor heterogeneity.
• Validated the pipeline's performance across different cancer types and evaluation metrics.

Conclusions:

• The developed CNN-based pipeline effectively classifies histopathology images for various cancer diagnostic tasks.
• Deep learning approaches, particularly Google's Inceptions, show significant promise in enhancing precision pathology.
• The pipeline provides a robust tool for automated cancer diagnosis, available open-source for further research.