Artificial Intelligence Merges Histology, Genomics to Predict Survival

A deep learning computational approach can predict survival with brain tumors directly from digital histological images and genomic biomarkers, according to a study published online March 12 in the Proceedings of the National Academy of Sciences. The so-called survival convolutional neural networks (SCNNs) show accuracy that surpasses the current clinical paradigm for predicting the overall survival of patients diagnosed with glioma, the authors say.

Human assessments of histology, while an important tool in cancer diagnosis and prognostication, are highly subjective. Machine learning or artificial intelligence has emerged as an important image analysis tool for medical imaging.

The present study explains the SCNN approach, which combines deep learning with traditional survival models to identify survival-related patterns from histology images.

Hematoxylin and eosin stain-stained tissue sections are first digitized to whole-slide images. These images are reviewed using a web-based viewer to identify regions of interest that contain viable tumor with certain histologic characteristics. An image sampling and risk filtering technique significantly improves prediction accuracy by minimizing the effects of intratumoral heterogeneity. High-power fields are sampled from these regions of interest and are used to train the neural network to predict patient survival. (SCNNs recognize important structures, like microvascular proliferation, that are used by pathologists in grading and prognosis.) Predictions are compared with patient outcomes to adaptively train the network. 

The SCNN approach was validated by predicting overall survival in gliomas using data from the Cancer Genome Atlas. SCNN predictions were highly correlated with both molecular subtype and histological grade and were consistent with expected patient outcomes.

"SCNN can effectively discriminate outcomes within each molecular subtype, effectively performing digital histologic grading," write the authors led by Pooya Mobadersany, from Emory University in Atlanta, Ga. "Using visualization techniques to gain insights into SCNN prediction mechanisms, we found that SCNNs clearly recognize known and time-honored histologic predictors of poor prognosis and that SCNN predictions suggest prognostic relevance for histologic patterns with significance that is not currently appreciated by neuropathologists."

Takeaway: These results suggest a role for artificial intelligence in precision medicine that will expand the use of computational analysis in the field of pathology.