DoMore! - A project for cancer patients and society
To best treat cancer, we must first understand how the disease will develop. DoMore! is a five-year-long project researching cancer and developing new cancer diagnostics methods based on artificial intelligence. The research has enhanced our ability to predict a patient's cancer development and provide the proper treatment.
The DoMore! project was in 2016 selected as one of the Norwegian Research Council's Lighthouse projects to solve large societal challenges using cutting-edge technology. With the use of digital tools as deep learning and "big data" from vast reserves of hospital patient data, we have found patterns that reveal information about a patient's cancer and the likely outcome - the prognosis. The projects is in it's final stage, and several products are already in the pipeline. These are roughly divided into three main groups to fit into the different parts of the industry's value chain:
- Pathology workflow optimisation
The commercialisation process contributes to transforming the knowledge into commercial products for clinical use. A range of artificial intelligence products based on the DoMore!-research are being developed. The Technology Transfer Office (TTO) Inven2, in collaboration with ICGI, will be commercialising two of the candidates. At the end of 2021, five product candidates will be offered through the company DoMore Diagnostics AS.
DoMore Diagnostics has signed a service agreement with Unilabs, which provides laboratory services to hospitals in several countries using DoMore! 's method of Histotyping for colorectal cancer. In the longer term, the plan is to provide services directly to hospitals around the world.
The DoMore!-purpose has been to utilise new technology to reduce over-and undertreatment of cancer by providing a new system for cancer prognosis. The lack of objective and precise methods for cancer prognosis is the most crucial cause behind the prevalence of over-and under treatment of cancer patients worldwide.
Understanding the tumour
Predicting the development of the patient's cancer is complicated. A cancerous tumour can consist of areas with different abnormalities, and contains much more information than the human brain can handle. Some deviations may spread and be deadly. Others will remain calm. The heterogeneity of cancer has been a significant challenge in cancer treatment.
Calculations indicate that the pathologist's assessment of the severity of the cancerous tumour is correct in only around 60 % of the cases. Today only a portion of the tumour is sampled. The risk of ignoring the cells that can kill the patient is great. The analysis work is time-consuming and resource-intensive. In addition there is a worldwide shortage of specialists who can perform these tasks.
Understanding how a tumour will develop is essential for the proper treatment of cancer patients, but most cancerous tumours are heterogeneous - complex - and therefore, difficult to predict. Today's pathologists have to base prognostication of cancer on subjective assessments. To be on the safe side, many cancer patients get more treatment than they might need. Overtreated patients are susceptible to severe side effects and reduced quality of life. If the conclusion is incorrect, it may cause additional costs, undesirable side effects and, at worst, lead to the patient's death.
The extensive research within DoMore! is done in close collaboration with researchers from the University of Oxford's Institute for Cancer Medicine and Cancer Hospital and University College London (UCL), in addition to Norwegian contributors such as the University of Oslo (UiO) and Sykehuset i Vestfold (SiV).
The team consists of international experts in several scientific and medical fields, including digital image analysis, tumour pathology, cancer surgery, and oncology. Prof Håvard E. Danielsen at Institute for Cancer Genetics and Informatics leads the work.
Using big data and deep learning through neural networks, we can now map a tumour's heterogeneity. "Training" computers on existing patients data to recognise cancer, the researchers can get a better idea about the treatment options for the patient already at the time of diagnosis.
The DoMore!-team enables computers to analyse cancer samples by teaching the machine using thousands of samples from previous cancer patients. The process is by far faster and safer than the experts can perform. Initially, colon, prostate and lung cancer samples have been included in the work. Together, these forms of cancer account for 44 % of all cancer deaths. Simultaneously, work on several other cancer types has also started.
Today, four years after the startup, DoMore! represents a complete diagnostic system capable of predicting the outcome for the cancer patient based on data from previous patients.
With today's pathological methods, as many as 80 % of patients with colorectal cancer end up with an unclear prognosis. With the new AI-based system this figure is reduced to 12 %.
DoMore! provides new methods for cancer diagnostics based on artificial intelligence. A new, automated system for diagnostics is currently commercialised through Domore Diagnostics as.
Our researchers have been publishing several research articles, confirming that it is possible to predict cancer using deep learning and artificial intelligence.
By "training" computers to recognise cancer, we can now get a safer and faster prognosis based on existing patient data. New and groundbreaking methods will give the specialist a better basis for choosing the right treatment options for the patient already at diagnosis.
We have, in one of our studies, developed a biomarker of patient outcome after colorectal cancer. We did this by directly analysing scanned conventional haematoxylin and eosin-stained sections using deep learning.
How is it done?
The DoMore! research group has developed a clinically useful prognostic marker using deep learning allied to digital scanning of conventional haematoxylin and eosin-stained tumour tissue sections. The new method is named Histotyping or, more precisely, the DoMore-v1-CRC marker. The assay has been extensively evaluated in large, independent patient populations and has shown to correlate with and outperforms established molecular and morphological prognostic markers and provide consistent results across tumour and nodal stage.
The biomarker stratified stage II and III patients into sufficiently distinct prognostic groups. A tool like this has the potential to guide the selection of adjuvant treatment by avoiding therapy in very low risk-groups and at the same time identify patients who would benefit from more intensive treatment regimes.
We have used more than 12 000 000 image tiles from patients with a distinctly good or poor disease outcome from four cohorts. These were used to train a total of ten convolutional neural networks, purpose-built for classifying supersized heterogeneous images. A prognostic biomarker integrating the ten networks was determined using patients with a non-distinct outcome. The researchers tested 920 patients with slides prepared in the UK and then independently validated according to a predefined protocol in 1122 patients treated with single-agent capecitabine using slides prepared in Norway. All cohorts included only patients with resectable tumours and a formalin-fixed, paraffin-embedded tumour tissue block available for analysis. The primary outcome was cancer-specific survival.
828 patients from four cohorts had a distinct outcome and were used as a training cohort to obtain clear ground truth. 1645 patients had a non-distinct outcome and were used for tuning.
The biomarker provided a hazard ratio for poor versus good prognosis of 3·84 (95% CI 2·72–5·43; p<0·0001) in the primary analysis of the validation cohort. After adjusting for established prognostic markers significant in univariable analyses of the same cohort (pN stage, pT stage, lymphatic invasion, venous vascular invasion) the hazard ratio was 3·04 (2·07–4·47; p<0·0001).
Collectively and separately, the results presented in the publications show that it is possible to teach a computer through Deep Learning and Big Data. We can not only do the same as today's pathology. We are also able to establish a more robust grading system in cancer types where today's praxis is less successful while at the same time, eliminating the subjective component.
In February 2019, halfway into the 5-year timeline–the Norwegian government chose to launch the news of a national AI strategy during a visit to the research Institute for Cancer Genetics and
Informatics. Prime Minister Erna Solberg and then Minister of Digitalisation Nicolai Astrup emphasised DoMore! and the positive consequences AI-based prognosis will have for patients and society through faster and more accurate cancer prognostication. The DoMore-model's potential is described in Norway's National Health and Hospital Plan 2020-2023, and the Norwegian government has, on several occasions, referred to the research as an example of good public investment in AI-based innovation.