Combat: Cancer and Computer Science
GlioVision: Combating Cancer with Computer Science
Kavya Kopparapu
Thomas Jefferson High School for Science and Technology
This article was originally included in the 2018 print publication of Teknos Science Journal.
An occasional headache. Continued hip pain. Persistent neck soreness. Three unrelated symptoms, minor individually, combined to present a most devastating diagnosis: stage 4 cervical cancer with metastasis to the lungs and brain. This was the story of my close friend’s mother, whose world flipped overnight in November 2015. Throughout the ordeal, we drew comfort and security from the success of chemotherapy on a number of patients. There was no time for expensive gene testing and no definite way for doctors to know whether a certain treatment would work. Over the weeks, we prayed that the medicine administered would be effective. But by Christmas, she had passed away.
Unfortunately, her story is not unique; aggressive cancers continue to impact families across the world. So how can we give both patients and doctors confidence that treatments will work? How can we shorten the current treatment pipeline and present the most targeted and effective treatment? How can we ease cancer’s impossible burden?
The answer lies in developing methods to determine personalized cancer treatments quickly and accurately. Only in recent years have treatments become tailored to a patient’s specific characteristics; previously, the only therapies were for an “average” patient. The issue was the basis of the 2015 Precision Medicine Initiative, which focused mainly on treatments for cancer and genetic diseases [7]. Through this program, we have made leaps and bounds towards creating new types of chemo- and immunotherapies targeted for a cancer’s specific genetic mutations.
However, a major disconnect exists in the current personalized treatment pipeline for aggressive cancers. The current procedure is as follows: patients are first diagnosed with cancer; then, information about the tumor is extracted using a combination of radiology, pathology, and genetic studies; finally, a specialized treatment is prescribed. However, while the diagnosis and treatment portions of the pipeline have experienced major innovations, few breakthroughs have tackled the speed and accuracy at which tumor information is extracted. This disconnect is likely why mortality rates for patients with aggressive cancers have stagnated in the past thirty years [3].
Individuals with glioblastoma, the most aggressive type of brain cancer, are an example of patients in grave need of bridging this divide: they have an average survival of less than one year post-diagnosis [1]. Obtaining accurate tumor information in a timely manner greatly increases the specificity of targeted treatment by an oncologist.
With the obvious need for a computational technique to determine this information, glioblastoma became the focus of my research at Stony Brook University on Long Island. Stony Brook is a center of biomedical informatics research, with projects ranging from assisting physicians to visualize medical images by rendering them in virtual reality to computational pathology projects aimed to diagnose diseases. After studying the biological basis for glioblastoma, I developed a computational platform, GlioVision, which extracts important tumor information for pathologists and oncologists.
GlioVision generates the following information about a glioblastoma tumor from a brain biopsy slide using deep learning: annotated areas of relevant tissues, molecular subtype, and expression status of an important gene called MGMT. It has achieved 86%, 94%, and 96% accuracy in these three tasks, respectively. These three tumor features—tissue locations, molecular subtype, and MGMT expression status—are significant in prognosis and treatment. Current methods take over one week to gather this information from expensive gene tests, while GlioVision’s data-driven approach takes less than five seconds for a fraction of the cost.
The GlioVision platform is unprecedented with its low cost, high speed, and improved accuracy. Since numerous previous studies have diagnosed glioblastoma using computational techniques with high accuracy, this problem was not the focus of GlioVision [2]. While previous studies have segmented pertinent regions of interest, none have isolated the same number of diagnostically-relevant regions that GlioVision has [5]. GlioVision has surpassed the accuracy of previous MGMT and gene methylation prediction studies [6]. The use of computer algorithms to determine GBM molecular subtype has not previously been attempted in published literature. Fusheng Wang, head of a biomedical informatics lab at Stony Brook University indicated that, “This work has two major contributions on digital pathology for glioblastoma. It developed an advanced machine learning based approach for quantitative analysis of digital pathology images through state-of-the-art deep learning techniques. In addition, the system performs integrative analysis with image features and molecular data, which has potential to advance glioblastoma research with precision medicine” (F. Wang, personal communication, January, 2018).
This is the first year a project like GlioVision could be implemented in hospitals. With the April 2017 FDA approval of slide-scanning technology, pathologists can now make medical diagnoses from scanned microscope slides on a computer. In order for GlioVision to be implemented, more hospitals need to adopt digitized microscope slides in their diagnosis pipelines.
While GlioVision currently targets brain cancer, the idea of extracting tumor information from a histopathological image is universal. GlioVision sets the proof-of-concept precedent for further research on extracting genetic and molecular information from a pathology image. The potential impact of this platform is tremendous; patients all over the world diagnosed with cancer could have access to a personalized cancer treatment regimen while avoiding costly genetic tests, easing the financial burden placed on families. They would also have access to this treatment quicker and, therefore, have a better chance of beating their cancer.
While presenting GlioVision at a conference, a fellow researcher explained that a very close friend of his was currently undergoing treatment for aggressive non-small cell lung cancer. His friend waited over a month for genetic test results locating the specific mutations her type of cancer possessed so that doctors could prescribe gene therapy. Anecdotes from families touched by cancer through a loved one are my motivation to continue to develop GlioVision into systems for more types of cancer.
I have tackled the most aggressive form of cancer, glioblastoma, in hopes that patients have access to the most effective treatment to target their specific tumor. I have tackled this project in hopes that one day, patients will get the best treatment for their specific cancer. I have tackled the divide in personal medicine so that in the future, families will be armed with a more better options in cancer treatment.
References
[1] American brain tumor association: Glioblastoma and malignant astrocytoma [Pamphlet]. (2017). Retrieved from http://www.abta.org/secure/glioblastoma-brochure.pdf
[2] Brain tumor diagnosis using deep neural network (DNN). (2016). International Journal of Innovative Research in Computer and Communication Engineering. https://www.ijircce.com/upload/2017/may/332_BRAIN_IEEE.pdf
[3] Center for neuro-oncology: How we diagnose brain tumors. (n.d.). Retrieved from Dana Farber Cancer Institute website: http://www.dana-farber.org/brain-tumors/diagnosis/
[4] FDA allows marketing of first whole slide imaging system for digital pathology. (2017, April 12). Retrieved from Food and Drug Administration website: https://www.fda.gov/newsevents/newsroom/pressannouncements/ucm552742.htm
[5] Hou, L., Samaras, D., Kurc, T., Gao, Y., Davis, J., & Saltz, J. (2015). Patch-based convolutional neural network for whole slide tissue image classification. Cornell University Library arXiv. Abstract retrieved from https://arxiv.org/pdf/1504.07947.pdf
[6] Korfiatis, P., & Klein, T. (2016). MRI texture features as biomarkers to predict MGMT methylation status in glioblastomas. NIH PMC. Retrieved from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4866963/
[7] The precision medicine initiative. (2015). Retrieved from The White House: President Obama website: https://obamawhitehouse.archives.gov/node/333101