Scientists at Imperial College London and the University of Melbourne have developed artificial intelligence (AI) software that they claim can more accurately predict ovarian cancer prognosis than current methods, and which can also indicate which treatments are most likely to be effective for individual patients.
The software, which has been trialed through a study at Hammersmith Hospital in the U.K., could pave the way to more effective, personalized treatment, suggested Eric Aboagye, PhD, professor of cancer pharmacology and molecular imaging at Imperial College London, and who is corresponding author on the team’s published paper in Nature Communications. “Our technology is able to give clinicians more detailed and accurate information on how patients are likely to respond to different treatments, which could enable them to make better and more targeted treatment decisions.” Hammersmith Hospital is part of the Imperial College Healthcare NHS Trust.
The researchers reported on their trial with the new software in a paper titled, “A mathematical-descriptor of tumor-mesoscopic-structure from computed-tomography images annotates prognostic-and molecular-phenotypes of epithelial ovarian cancer.”
Epithelial ovarian cancer (EOC) is the sixth most common cancer in women in the U.K., and has the highest mortality rate of all gynecological cancers, the authors stated. The five-year survival rate for EOC is about 35–40%, and the cancer accounts for 4% of all cancer deaths in women. Diagnosis is commonly not made until symptoms become evident at a later, rather than early stage.
High-grade serous ovarian cancer (HGSOC) is the most prevalent and most deadly subtype of the disease, affecting about 70% of EOC patients. Genomic profiling studies have uncovered a number of putative prognostic biomarkers of chemotherapy resistant and refractory HGSOC, while microRNA data has also been used to help stratify EOC risk. However, “it remains challenging,” the authors wrote, “to translate these molecularly determined characteristics into clinically relevant biomarkers due to intra-tumor heterogeneity, additional high assay cost, and time delays.” What’s needed is a real-time, noninvasive and cost-effective prognostic marker approach to help identify the best treatment for individual EOC patients.
Ovarian cancer diagnosis may typically involve a blood assay to test for the biomarker CA125, and CT scans to provide a detailed view of any tumors. The tumor imaging can help to identify how far the disease has progressed and spread, and so indicate whether surgery may be a treatment option in addition to chemotherapy. But what scans can’t do is predict overall outcomes or how effective the different therapeutic options will be in reality.
“The long-term survival rates for patients with advanced ovarian cancer are poor despite the advancements made in cancer treatments,” Aboagye said. “There is an urgent need to find new ways to treat the disease.”
The researchers’ new computational tool, called TEXLab, uses machine learning to evaluate ovarian tumors based on a detailed analysis of four biological characteristics—structure, shape, size, genetic makeup—which impact on overall survival. The results are generated as a Radiomic Prognostic Vector (RPV) score, which indicates the severity of the disease and likely prognosis.
In the reported trial the software was tested to 657 different quantitative mathematical descriptors, on preoperative CT scans from 364 EOC patients. Protein expression and genomic profiles of a subset of patients were also evaluated, and fresh frozen tissue samples from primary EOC patients were analyzed so that histological, protein, and gene expression data were all available to provide context alongside the RPV results.
The study results indicated that the RPV score was up to four times more accurate at predicting ovarian cancer death than current blood test and other prognostic scores. The data also suggested that 5% of patients with high RPV scores had a survival rate of less than 2 years, while a high RPV result was similarly linked with tumor resistance to chemotherapy and poor surgical outcomes.
“Notably, RPV possessed a better prognostic power when compared to the existing prognostic markers including CA125 and the transcriptome-based molecular subtype and potentially synergizes with existing CT-based morphological approaches,” the authors wrote. “We demonstrate, based on the strong association between RPV and response to primary chemotherapy or surgery, that patients with high RPV have a significantly high risk of failing quality surgery or systemic strategies and suggest that they possibly need to be directed towards alternative therapeutic approaches.”
These overall data indicated that the RPV could represent a biomarker for predicting treatment response, and might be used to identify patients who are less likely to benefit from standard therapy and who may require alternative treatment approaches. “Artificial intelligence has the potential to transform the way healthcare is delivered and improve patient outcomes,” commented study co-author Andrea Rockall, PhD, honorary consultant radiologist at Imperial College Healthcare NHS Trust. “Our software is an example of this and we hope that it can be used as a tool to help clinicians with how to best manage and treat patients with ovarian cancer.”
Interestingly, assessing the RPV data in the context of proteomic, transcriptomic, and genetic analyses of the tumors indicated that high RPV scores were associated with particular stromal phenotypes and activation of DNA damage response pathways, which could feasibly hint at therapeutic targets. “Overall, stromal phenotype on one hand, and proliferation and DNA damage response pathways on the other, were respectively activated in RPV-high and RPV-low tumors, all of which are potential actionable therapeutic targets in HGSOC,” the team commented.
“In the present study, we obtained and analyzed a comprehensive radiomic profile containing features for 364 EOC cases in total—the largest study of its kind for EOC,” the team summarized. “We discovered a novel radiomics-based prognostic signature, RPV, that not only has strong prognostic power (HR> 3), but is also noninvasive and readily accessible, compared to the existing molecular profiles and clinical factors deemed prognostically relevant.”
A larger study is now being planned to find out just how accurately the software can predict surgical outcomes, and how individual patients will respond to particular treatments.
Importantly, the RPV prognostic model is simple, and only requires information from a patient’s routine CT scan, which is available immediately without extra costs. The software is also fast, and can compute the RPV of 80 EOC datasets within 5 minutes on a standard computer, the researchers pointed out.
While the predictive power of the RPV approach was validated in two independent cohorts of patients, the researchers acknowledge that their study was retrospective in design, and a future prospective study or analysis of retrospective randomized clinical trial data should be carried out. Nevertheless, they concluded, “… we have discovered and validated a novel mathematical descriptor of tumor phenotype and prognosis that convincingly fulfills an unmet need in the management of patients with EOC, and have demonstrated a disruptive technology that opens the way for multiple classifications of patients and rapid patient entry into clinical trials at the point of care.”