Crowdsourcing Innovation in Cardiovascular Imaging
Friday, June 03, 2016
The ACC recently partnered with Booz Allen Hamilton and the National Heart, Lung, and Blood Institute (NHLBI), to put data science to work in the cardiology field in the 2015 Data Science Bowl – a competition that uses data science for social good. This year’s competition challenged data scientists to create an algorithm to automate the process of assessing cardiac function via magnetic resonance imaging (MRI). The National Institutes of Health and Children’s National Medical Center compiled data from more than 1,000 patients for examination – a data set an order of magnitude larger than any previously released data set of its kind. With it came the opportunity for the data science community to take action to transform how clinicians diagnose cardiovascular disease.
Prem Soman, MD, FACC, immediate past chair of ACC’s Imaging Section, and Victor Ferrari, MD, FACC, current chair of the Imaging Section and Immediate Past President of SCMR, sat down with Cardiology to explain the impact of the results of the Data Science Bowl. For further information, visit DataScienceBowl.com.
What kind of impact would automated, real-time MRI results have on cardiovascular imaging?
PS: While MRI is considered to be highly repeatable for left ventricular function assessment, the requirement for manual border tracing does introduce some operator dependence. The ability to automate the process will therefore probably improve precision. This is assuming that the approach to automation and the algorithm used are robust and do not affect precision adversely.
VF: One of the real-world barriers to adoption of newer imaging techniques is the time required for data analysis. Automated, real-time cardiac magnetic resonance analysis would be tremendously helpful in increasing efficiency and throughput for patients, and greatly reduce the total time from the scan to final report. If analysis algorithms could run automatically in the background following a scan, it would also shorten the time to obtain quantitative data, and increase exam efficiency.
If the NHLBI finds the winning algorithms to be effective, do you anticipate that these solutions will be readily adopted into practice?
PS: If such an approach is shown to be superior to manual border tracing, it is likely to be accepted in practice, particularly if it results in less operator time as well.
VF: Yes, with some modifications. The more human interaction needed to prepare the data, the greater the overall inefficiency. If we can develop robust feature-tracking algorithms that can accurately and more fully automate data analysis, they will quickly make their way into practice. The capability to run analysis programs in the background while other data are being acquired will increase exam value and prove more attractive to users.
How can competitions like the Data Science Bowl help to bridge the gap between proving the effectiveness of crowdsourced algorithms and actually implementing solutions in clinical practice?
PS: The power of crowdsourcing is tremendous and has been harnessed effectively by entities such as Wikipedia. Competitions such as the Data Science Bowl could be an effective mechanism to do the same in medicine.
VF: Increasing the productive interactions between the medical and data science communities will go far to bring us clinically relevant automated data analysis and other tools. Events such as the Data Science Bowl raise the visibility of the current clinical needs from a software standpoint, and allow us to better define robust solutions using state-of-the-art techniques.
What do you see as the future of cardiovascular imaging?
PS: Imaging is a critical component of cardiovascular medicine. There are few cardiology patient encounters that are completely devoid of an imaging component. Thus, the value of imaging service is obvious. However, as paradigms of health care delivery change, and value-based utilizations become the norms, it will be imperative for us to demonstrate the value of imaging in patient care more objectively, in terms of cardiovascular outcomes.
VF: Two major areas will have increasing relevance – 1) analyzing large datasets to answer critical important questions, such as the Hypertrophic Cardiomyopathy Registry and the Global Cardiovascular Magnetic Resonance Registry, and 2) demonstrating which cardiovascular imaging tests have the greatest value (clinical impact per dollar spent).
Do you see artificial intelligence playing a larger role in imaging?
PS: Yes, there already are major efforts to harness the strengths of machine learning in clinical imaging applications. Analyses of “big data” have provided unique insights – this trend can only get stronger.
VF: Definitely – moving away from human-based to machine-based analysis will be vital to progress in imaging. Pattern recognition in image analysis is important for disease detection, and newer "imaging intelligent agents" will help us identify patterns that the human eye and brain may not be capable of. They may well be capable of accumulating and interpreting data in order to propose diagnoses at a speed well beyond that of humans.
I believe that the Data Science Bowl was one of the most effective demonstrations of how a novel algorithm can work toward a feasible solution that has great clinical impact. With greater sophistication, "plug and play" algorithms that are fully automated – and reliable and accurate – will be a major advance in clinical quantitative cardiac imaging.
Article provided by the American College of Cardiology.