Sizing Up Technology Symbiosis
Specialists who embrace these new developments have a bigger toolbox than ever.
Machine learning (ML) is on the rise in just about every field of medicine, signaling changes that have some specialists speculating on how the ever-improving technology may change their position in the health care landscape. Radiologists may feel particularly unsteady as computer-aided detection (CAD) and diagnostic algorithms produce impressive results that test the mettle of human counterparts.
A good example in radiology is mammography CAD, which is now used in approximately 90 percent of mammography practices, according to Eliot L. Siegel, MD, FACR, professor of radiology at the University of Maryland School of Medicine. By analyzing images for suspicious traits, most experts and the ACR agree, CAD can be an invaluable screening tool.
A CAD algorithm has been able to detect and characterize various types of spinal metastases on PET/CT images with great accuracy. According to the authors of a recent study, the automated CAD system could help physicians more easily diagnose and treat patients with the most common kinds of musculoskeletal malignant growths. In medical oncology and cardiology, algorithms and machine-learning risk decision trees are used to guide treatment decisions and provide recommendations related to eating habits and exercise.
“Radiology is far ahead of other specialties in its research on machine learning,” Siegel says. “We have been conducting research in this area for more than 30 years.”
ML in Other Specialties
For radiology, the inherent value in machine-learning algorithms that can consume huge sets of data to process millions of images and medical observations is clear. Other specialties are also considering new technology in search of better patient outcomes, rather than treating ML improvements as a threat.
Urologists, for example, are exploring the use of ML algorithms to predict the recurrence of bladder cancer. Researchers have used the genes in a molecular signature to predict a patient’s five-year risk of recurrence after the resection of a bladder tumor. Their findings could aid in the management of non-muscle invasive urothelial carcinoma. Mobile ultrasound bladder screening devices now offer more imaging capabilities than traditional bladder scanners and create an opportunity for machine learning through the collection of patient diagnostic data.
Machine-learning algorithms last year were credited with distinguishing between pathological hypertrophic cardiomyopathy (HCM) and physiological changes in echocardiographic images. Partho Sengupta, MD, director of cardiac ultrasound research and professor of medicine in cardiology at the Icahn School of Medicine at Mount Sinai, said in a press release, “Machine-learning algorithms could assist in the discrimination of pathological versus physiological hypertrophic remodeling, … enabling easier and more accurate diagnosis of HCM.”
Integrative patient-focused machine-learning models developed in the Precise Medical Diagnostics program at the Icahn School of Medicine at Mount Sinai’s Center for Computational and Systems Pathology have been designed to demonstrate the importance of histology, clinical features, and biomarker profiles. These profiles are used to drive more effective treatment strategies. “Such standardized approaches will position diagnostic and predictive pathology to be at the forefront of effective patient management,” says Carlos Cordon-Cardo, MD, PhD, who heads the department of pathology at Mount Sinai. “Creating objective and standardized methods for assessing biomarker relevance in conjunction with clinical variables and tumor pathology was the basis for some of the early prostate cancer models that have helped to guide treatment decisions for men newly diagnosed with prostate cancer.”
“Machine learning is a tool to improve efficiency, diagnostic accuracy, and safety,” Siegel says. “With the arrival of these tools over the next many years, we should view them as exciting and wonderful enhancements to our practices.”
With machine learning becoming more mainstream and affordable — and boasting high rates of accuracy and probability of a correct diagnosis — its evolution could benefit smaller practices coping with rising costs and staffing shortages. As radiological services are enhanced, there still will be a need for personalized, human recommendations in imaging as they relate to further testing and treatment.
Pairing ML algorithms with automated scanning capabilities, for example, allows computers to look for undiscovered disease while a radiologist focuses on her diagnostic target. CAD uses a dataset to process an image and signal areas where disease might be present. However, even with CAD, there’s typically a parallel track where a radiologist also interprets images and then decides what’s true or false.
“It is fairly well recognized that in medical applications there are tasks that are better suited to computational approaches and other tasks that are better suited to human approaches,” says Cordon-Cardo. “Humans are much more capable at detecting model failures and outliers, as well as reconciling conflicting information, whereas computers are much better at keeping track of relationships and building models off of these relationships. Success in this field will therefore always be measured in terms of the ability to efficiently combine and integrate the strengths of both.”
“Essentially, there is no need to panic,” Siegel says. “We are not even close to seeing the kind of ‘general artificial intelligence’ that would be required to even contemplate replacing a radiologist.” He says that when such technology is one day ready to use, it could take 20 years beyond that point to fully test its capabilities and to get FDA clearance for what would be an incredibly comprehensive system.
Until that happens, greater education is key, says Cordon-Cardo. “Spend time understanding how specific machine-learning models are constructed,” he says. “Study the composition of the patient group used to define the model and the features and attributes that drive a model’s accuracy.” Specialists should strive to understand these new tools as they become available, he says, and be able to use them to improve quality and speed. “Mastery of this toolbox will be valued and provide increased differentiation capability for specialists.”
“We are on the cusp of being the first specialty to embrace ML as an effective partner and friend,” says Siegel. “I’m certain computers will not replace human evaluation and diagnosis in the practice of anyone reading this, including those of you still in training today.”
By Chad Hudnall, writer for the ACR Bulletin