Predicting cancer outcomes from histology and genomics using convolutional networks

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Precision medicine

Background:

• Cancer histology provides rich phenotypic information predictive of patient outcomes.
• Current clinical paradigms for survival prediction have limitations.

Purpose of the Study:

• To develop a computational approach for learning patient outcomes from digital pathology images.
• To integrate histology and genomic data for improved survival prediction.

Main Methods:

• Utilized deep learning, specifically survival convolutional neural networks (SCNNs).
• Employed statistical sampling techniques to address tumor heterogeneity and training data needs.
• Integrated histology images and genomic biomarkers into a unified framework.

Main Results:

• SCNNs demonstrated prediction accuracy surpassing current clinical paradigms for glioma overall survival.
• Heat map visualization revealed SCNNs recognize prognostically important histological structures like microvascular proliferation.
• The approach successfully integrated histology and genomic data.

Conclusions:

• Deep learning, via SCNNs, offers a powerful tool for precision medicine in oncology.
• Computational analysis of histology shows expanding utility in pathology practice for outcome prediction.