Deep Learning–Based Tissue Analysis May Benefit Colorectal Cancer Patients

Deep learning techniques may pave the way for a more accurate outcome prediction in colorectal cancer patients as compared to evaluations currently performed by an experienced human observer. Researchers at the University of Helsinki are reporting in Scientific Reports that they have created a deep learning algorithm that appears to help clinicians better predict patient outcomes based on colorectal cancer tissue samples.

“In our study we hypothesized whether a deep learning–based algorithm can be trained to extract prognostic features from cancer tissue images without any expert-defined supervision. It appeared exciting that almost no domain expertise is needed to build accurate classifiers,” said study investigator Johan Lundin, MD, PhD, who is the Research Director of FIMM-Institute for Molecular Medicine Finland, at the University of Helsinki.

The researchers combined convolutional and recurrent architectures to train a deep network to predict colorectal cancer outcome based simply on images of tumor tissue samples. The researchers noted that their latest findings are novel because this algorithm was able to predict patient outcome without any intermediate tissue classification.

The team evaluated a set of digitized hematoxylin-eosin stained tumor tissue microarray (TMA) samples from 420 patients with colorectal cancer. The researchers had clinical, pathological, and outcome data on all of the patients. The deep learning–based outcome prediction was able to outperform two other methods with only small tissue areas as input. The researchers found it was more accurate than visual histological assessment performed by human experts on both TMA spot and whole-slide levels. 

“Training a machine-learning classifier supervised by patient outcome instead of expert-defined entities has the potential to identify previously unknown prognostic features. Our results indeed suggest that deep learning techniques enable a more accurate outcome prediction as compared to an experienced human observer,” Lundin explained.

The five-year survival outcome prediction by the algorithm outperformed the assessment by human experts. The same set of images was shown to three experienced pathologists from two different institutions. When the researchers compared the performance of the automated analysis against that of the pathologists, they were pleased to see the results. The machine learning–based approach did a better job than the human observers in categorizing patients into long-term and short-term survivors. In determining histological grade, the learning-based approach outperformed the conventional microscopy analysis of the whole-slide tumor sample.

Dr. Lundin said further research is warranted to better understand what factors affect the final decision of the classifier and which features drive the predictions. The researchers hope to better define what the neural network “sees.” Dr. Lundin emphasized that these initial findings are an important step toward delivering more personalized medicine for colorectal cancer patients. “Outcome prediction is crucial for patient stratification and disease subtyping in clinics, to aid in decision-making and eventually deliver more personalized treatments for cancer patients,” he said.