Novel AI Algorithm Provides Prognosis for Advanced Ovarian Cancer

Becky McCall

February 18, 2019

A novel algorithm that utilises a non-invasive measure, known as the Radiomic Prognostic Vector (RPV), has been developed and validated to help determine the prognosis and response to treatment in advanced ovarian cancer.

By determining subtle differences in the phenotype of epithelial ovarian tumours, data drawn from computed tomography (CT) scans are fed into a mathematical software tool called TexLab to generate an RPV score. This score indicates how severe the patient's disease is, ranging from mild to severe, and provides a measure of prognosis according to how aggressive the tumour is.

It is the first time such a tool has been developed for this purpose. "It could be transformational for these patients," study lead Eric Aboagye, professor of cancer pharmacology and molecular imaging at Imperial College London, told Medscape News UK. The work was carried out in collaboration with the University of Melbourne, Australia, and is published in the February 15th edition of Nature Communications.

Novel Tool

"The algorithm is a novel tool. It's not just better than existing tools but it is completely different and novel in terms of stratifying patients," emphasised Prof Aboagye. "We hope that it can be used in the future for personalised therapy of advanced stage ovarian cancer, and potentially, after some further work, for other cancers too."

He added that the tool demonstrated how a disruptive technology could open the way for multiple classifications of patients and could facilitate rapid patient entry into clinical trials at the point of care. "Every patient gets a CT scan anyway and we wanted to use this because it quick – it doesn't take 3 weeks for a genome sequence; it's inexpensive and non-invasive."

Predicting Prognosis May Improve Poor Outcomes in Advanced Ovarian Cancer

There are about 6000 new cases of ovarian cancer per year in the UK. The long-term (5 years) survival rate is just 35-40% because the disease is often diagnosed at a late stage, as symptoms such as bloating tend to be non-specific. 

Despite significant advances in cancer therapies, there remains an urgent need to find new ways to treat the disease. Recent advances in molecular and genomic profiling have resulted in significant prognostic biomarker discoveries, however, the translation of these molecular characteristics into clinically relevant biomarkers is challenging, write the authors. "There's also huge variation in outcome with maximal effort surgeries and chemotherapies with some patients surviving for 2 years, and some for 5," said Prof Aboagye.

"With the algorithm, we wanted to stratify patients into groups according to predicted survival to help determine the best treatment for each group," he explained. The algorithm's development used radiomics, a process that extracts a large amount of quantitative data from medical images using data-characterisation algorithms. In ovarian cancer, this involves quantifying the mesoscopic [between micro- and macroscopic] tumour phenotype from anatomic or functional images, and defining tumour spatial complexity – including first and higher order statistics, fractal and shape features, generating descriptors of disease features not visible to the naked eye.

Prof Aboagye explained how they analysed pre-surgical CT scans from 364 ovarian cancer patients for nearly 660 quantitative mathematical descriptors that could be important in cancer progression. Using artificial intelligence (AI), the researchers derived the RPV summary-statistic and its associated prognostic algorithm based on the four descriptors that most accurately predicted mortality. 

The descriptors relate to tumour macro-architecture, and in biological terms, the individual components of RPV combine to define tumour mesoscopic structure (structure, shape, size and genetic makeup):

a) maximal fractal dimension of the tumour and its microenvironment, which was negatively correlated with survival, together with the following positively correlated features;
b) proportions of runs revealing coarse low density textures, e.g. intermixed fibrotic stroma and tumour cells; 
c) the average visual contrast across the tumour, and
d) a measure of the global heterogeneity of the entire tumour. 

The researchers compared the RPV results with blood tests and current prognostic scores used to estimate survival. They found that the software was up to four times more accurate for predicting deaths from ovarian cancer than standard methods. They found that the 5% of patients with high RPV scores had a median survival rate of less than 2 years. High RPV was also associated with chemotherapy resistance, shorter progression-free survival and poor surgical outcomes.

Prof Aboagye pointed out that the research had also provided insight on why the difference in survival times might come about. "We see the stroma and DNA damage response as mechanisms that might explain why the algorithm manages to stratify patients so well. This will lead to changes in how we treat patients."

He added that in terms of immediate use of the algorithm to stratify patients for optimal treatment, this is currently only possible for standard surgery plus chemotherapy. "We are close to being able to provide the right treatment for the right patient with maximal effort cytoreductive surgery, but with other specific therapies, more work is needed. For example, for DNA damage response there are currently clinical trials exploring poly ADP ribose polymerase (PARP) inhibitors, and with a little more work we should arrive at even more specific predictive vectors for this particular patient group."

Next Steps

Prof Aboagye said that they intend to carry out a larger prospective study to see how accurately the software can predict the outcomes of surgery and/or drug therapies for individual patients.

He added that the current study focused on advanced, lethal, high-grade tumours but that they hoped to expand this to all tumours eventually. "One of the challenges we face is the availability of CT scans and related clinical data for all the cancers we'd like to develop algorithms for." With a view to using the software for other tumour types, they have developed a gene expression profile as a surrogate marker for RPV, which could potentially be used to assess other cancers.

The algorithm runs on a standard personal computer and can process 80 algorithms in 5 minutes, using minimal computing power. Furthermore, very little extra resource and expenditure is required to use the software.

COI: One author, Hani Gabra is an employee of AstraZeneca. The other authors declare no competing interests.

Published the February 15th  edition of Nature Communications


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.