How is machine learning (ML) impacting the field of medical imaging today, and what does it have to offer to the field of diagnostic medicine?
Medical imaging and diagnostics is a vital field of medicine, which includes many imaging modalities used to image parts of the human body to provide diagnoses and treatments of disease. Techniques using MRI, (PET)-CT/MRI, and 3D ultrasound imaging are some of the many tools radiologists use to diagnose illnesses ranging from cancer to Alzheimer’s.
The Government of Canada estimates that, as of 2014, Canada has a medical device market accounting for around 2% of the global market, worth an estimated US$6.7 billion. They also estimate that diagnostic imaging accounts for approximately 20% of the money spent as a percentage of total medical device sales in 2014 in Canada.
The impact Canada is making in this industry is significant. According to the Government of Canada, their innovation agenda “aims to make Canada a global centre for innovation.” Ontario is home to a number of world-class research centres specializing in medical imaging including Sunnybrook Research Institute, the University Health Network (UHN), and Robarts Research Institute, to name a few. According to a report released by MaRs in 2009, the University of Toronto has the largest Department of Medical Imaging in Canada.
There are many concerns about how advances in automation, specifically in ML, are changing the job landscape. Many argue that ML will lead to the “singularity,” a term coined by John von Neumann to describe a future point in human history beyond which civilization will be forever changed by rapid advances in technology. Many predictions on how machine intelligence will change society are apocalyptic, describing a path where humanity has been entirely replaced by machines in the workforce. To put these concepts into focus, let us see what role ML has in medical imaging today.
Nick Bryan, M.D., Ph.D., writing for the Radiological Society of North America, posed a question to explain the motivation behind employing ML in the medical industry: “Can a machine learn more than what we now know and use this knowledge to make decisions?” He uses the definition made by Arthur Samuel in 1959, describing ML as “the field of study that gives computers the ability to learn without being explicitly programmed.” According to Dr. Bryan, the decision-making tasks in radiology are mainly those of classification; a radiologist’s main task is to find the most likely diagnosis, given the available images and clinical information.
ML uses algorithms to make decisions. Some strategies include random forest classification, Bayesian networking, and genetic algorithms. No matter what the strategy, however, Dr. Bryant argues that an algorithm’s value to the radiologist depends on how accurately it makes a diagnosis. Since radiologists are familiar with statistical metrics, the outcomes of algorithm performance can be easily understood.
Dr. Bryan describes two general paradigms for designing ML algorithms: supervised and unsupervised learning. Supervised learning is a process where a radiologist teaches a machine, and the machine learns from what is known from inputs of imaging data with categorized outcomes. Alternatively, unsupervised learning is a process where the machine determines what the possible diagnoses are and how to discriminate against them by cycling through large data sets (ie. “bid data”). While the first method has the potential to increase the throughput of diagnoses, the second could potentially lead to new information on what patterns can be used to diagnose an illness without the help of human professionals.
The marriage between ML and medical imaging is becoming irreversible. In an editorial published in the Journal Pattern Recognition, Kenji Suzuki et al. describe how ML has become indispensable to the field of medical imaging. Difficulties due to variations in the complexity in biomedical image, or in deriving analytical solutions to represent objects like lesions and anatomies in biological imaging data, are some of the many seemingly insurmountable challenges facing the imaging and diagnostic profession. They describe problems in medical imaging as requiring “learning from examples/data for accurate representation of data and modeling of prior knowledge, which is exactly the focus of machine learning.”
A study by A. Alansary et al. leverage ML to tackle the challenges of imaging a fetus due to the variability in position and orientation. They extract superpixels, a segment of an image which is in better alignment with intensity edges than a rectangular patch, to construct a graph. They trained a random forest classifier to tell the difference between brain and non-brain superpixels. The method validation achieved a 94.55% accuracy rate of brain detection.
Dr. Bryan argues that when the success rate of the machine exceeds that of the human, it will have learned more than what a radiologist knows, giving it the ability to make decisions a human could not make. According to Bryan, this is not, however, the end of radiologists. Though machines are rapidly becoming capable of learning complex sets of data from large normal and diseased populations, he predicts that machines are destined to complement our human skills of pattern recognition.
Time will tell the what the long-term effect will be introducing machine learning into medicine. The capacity, however, of machine learning to advance the field of medical imaging is clear, and in a rapidly expanding industry it is likely to become an integral tool in the field of diagnostic medicine.