Home - Emerging Tech - Deep Learning Algorithm Applied to Tumor Tissue Can Predict Patient Outcome in Colorectal Cancer
Research Brief

Deep Learning Algorithm Applied to Tumor Tissue Can Predict Patient Outcome in Colorectal Cancer

Researchers have developed a deep learning model that directly predicts the probability of five-year survival in patients with colorectal cancer

Meeri Kim, Contributor
Thursday, March 15, 2018


Colorectal cancer diagnostics currently involves microscopic assessment of tissue morphology by a pathologist along with assessment of tumor stage. As a way to automate this process, several studies have proposed machine learning methods that can accurately classify features in medical images.

A team of researchers from the University of Helsinki in Finland has trained a deep learning algorithm that goes a step further by directly predicting patient outcome based on tissue samples of colorectal cancer. The technique does not include any intermediate tissue classification but still outperformed pathologists in stratification of patients into low-risk and high-risk categories. The study was published Feb. 21 in Scientific Reports.

The deep learning model was developed and trained to output a probability of survival at five years after primary colorectal cancer diagnosis. To test its prediction ability, the authors obtained digitized microscopic images of tumor tissue from 420 patients diagnosed with colorectal cancer with available clinicopathological data and known disease-specific outcome.

Patients were stratified into low- and high-risk groups individually based on three methods: the authors' machine learning model with an input of small sections of tumor tissue cores, a Visual Risk Score given by three experienced pathologists observing the same small sections of tumor tissue cores, and a histological grade assessed by conventional microscopy analysis of the whole-slide tumor sample.

The machine learning model predicted five-year disease-specific survival with higher accuracy than both the histological grade and the Visual Risk Score. In other words, for the task of stratifying into low- and high-risk patients, it was able to outperform human experts. The results of the study suggest that a deep learning algorithm shows promise and can extract prognostic information from even a small tissue area.