Exploring Histological Similarities Across Cancers From a Deep Learning Perspective.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital histopathology

Background:

• Histopathology image analysis is crucial for cancer diagnosis.
• The Cancer Genome Atlas (TCGA) provides extensive whole slide images across various organs and subtypes.
• Limited research has explored similarities and cross-organ applications of histopathology models.

Purpose of the Study:

• To train deep learning models for classifying cancer versus normal tissue patches across 11 subtypes and 7 organs.
• To investigate the cross-organ inference capabilities of these trained models.
• To explore the underlying reasons for observed cross-organ performance, hypothesizing shared tumor morphologies.

Main Methods:

• Developed deep learning models to classify cancer vs. normal patches from 9,792 whole slide images (11 subtypes, 7 organs).
• Evaluated model performance on test sets from different organs (cross-organ inference).
• Validated findings using high-dimensional feature space separability, Gradient-weighted Class Activation Mapping (GradCAM), and nuclei feature distribution analysis.

Main Results:

• Models demonstrated good cross-organ inference accuracy on breast, colorectal, and liver cancers.
• High accuracy was observed for models trained on cancer subtypes from the same organ (kidney, lung).
• Shared tumor morphologies across organs were identified as a likely cause for high cross-organ inference, supported by GradCAM and nuclei feature analysis.

Conclusions:

• Deep learning models trained on histopathology images exhibit significant cross-organ inference capabilities.
• Shared morphological features contribute to the generalization of these models across different organs.
• This study provides insights into the commonalities of cancer at the histopathological level, paving the way for broader AI applications in pathology.