FIMM Monthly Publication Highlight
February 2018: Deep learning based tissue analysis predicts outcome in colorectal cancer
Every month we publish a publication highlight that summarises some of the research taking place at FIMM.
A research team led by Research Director Johan Lundin has shown that deep learning techniques can predict colorectal cancer outcome based on images of cancer tissue samples, without intermediate tissue classification steps. The five-year survival outcome prediction done by the algorithm outperformed the assessment done by human experts.
The results of the work were recently published in Scientific Reports.
The main goal of the study was to investigate whether a deep learning algorithm that takes images of small regions of tumour tissue as input can be trained to predict outcome of cancer patients without prior knowledge of the disease or expert guidance.
The research team utilised digitally scanned tissue samples from 420 colorectal cancer patients with related disease outcome data.
With this dataset, the team developed and trained a machine learning model to directly predict the five-year disease-specific outcome with only one small tissue image per patient as input. Two deep learning methods, called convolutional and recurrent neural networks, were applied.
The same set of images was then shown to three experienced pathologists from two different institutions. When the researchers then compared the performance of the automated analysis against that of the pathologists they observed that the machine learning based approach outperformed the human observers in categorisation of patients into long-term and short-term survivors.
The machine learning-based method also outperformed histological grade, assessed based on conventional microscopy analysis of the whole-slide tumour sample.
We observed that our digital risk score is independent of both histological grade and stage of disease. This implies that even a small tissue section contains valuable information about the disease outcome and that artificial intelligence-based methods can be used to extract this information
- FIMM PhD student Dmitrii Bychkov, the first author of the article
Outcome prediction is crucial for patient stratification and disease subtyping to aid the clinical decision-making to achieve a more personalized treatment regimen.
Our hypothesis was that 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.
Further research is needed to understand what factors affect the final decision of the classifier and which features drive the predictions, i.e. what the neural networks “see”. The suggested model should also be trained on larger tumor tissue areas and evaluated on an extended patient series
- Research Director Johan Lundin, who led the study
Bychkov D, Linder N, Turkki R, Nordling S, Kovanen PE, Verrill C, Walliander
M, Lundin M, Haglund C, Lundin J. Deep learning based tissue analysis predicts outcome in colorectal cancer. Sci Rep. 2018 Feb 21;8(1):3395. doi: 10.1038/s41598-018-21758-3.