Deep learning-based cross-classifications reveal conserved spatial behaviors within tumor histological images

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Histopathological images are crucial for cancer diagnosis but manual analysis is time-consuming.
• Automated analysis of whole slide images (WSIs) using machine learning is needed for large-scale cancer research.
• Convolutional neural networks (CNNs) offer potential for analyzing complex image data.

Purpose of the Study:

• To develop and apply CNNs for analyzing histopathological images across various cancer types.
• To classify tumor/normal status, cancer subtypes, and mutations using CNNs.
• To explore conserved spatial behaviors and similarities across different tumor types.

Main Methods:

• Trained CNN architectures on 27,815 hematoxylin and eosin stained WSIs from The Cancer Genome Atlas (TCGA).
• Classified tumor/normal status, cancer subtypes, and TP53 mutations.
• Evaluated CNN performance using Area Under the Curve (AUC) metrics.
• Compared CNNs on pathologist-annotated nuclei in breast and colon cancer images.

Main Results:

• Achieved high AUCs (0.995 ± 0.008) for tumor/normal classification across 19 cancer types.
• Successfully classified cancer subtypes with significant accuracy (AUC 0.87 ± 0.1).
• Demonstrated cross-tissue transferability of CNNs (AUC 0.88 ± 0.11) and identified conserved spatial patterns.
• Detected TP53 mutations with AUCs ranging from 0.65-0.80.
• Found both cellular and intercellular regions contribute to CNN accuracy.

Conclusions:

• CNNs are powerful tools for histopathological classification and image data mining in oncology.
• CNNs can reveal conserved spatial behaviors and similarities across diverse cancer types.
• This approach facilitates large-scale analysis and comparative studies of tumor biology.