An expert's take on the future of machine learning in quantitative image analysis

Current and future clinical use and real-world applications of artificial intelligence (AI) technology in healthcare and medicine will take center stage when new research, startups, scientists and industry leaders come together from June 6 to 7 in Hong Kong for the AI in Healthcare Summit.  

Among the 30 speakers scheduled to present at the summit is W. Art Chaovalitwongse, PhD, a professor of industrial engineering, 21st Century Leadership Chair in engineering, and co-director of the Institute of Advanced Data Analytics at the University of Arkansas. His presentation, entitled "The Future of Machine Learning in Quantitative Medical Image Analysis," will expand upon his work comparing radiomics and deep learning-based features to predict clinical outcomes from medical imaging data. 

Over the course of his career, Chaovalitwongse has held various faculty positions in universities and has consulted for companies such as ExxonMobil and Cisco, in addition to writing and co-writing over 150 published papers in machine learning, data analytics and optimization with applications in healthcare and other fields. 

Health Imaging spoke with Chaovalitwongse about his work with machine learning and quantitative medical image analysis.

Electroencephalography, neuroimaging and feature generation engineering  

According to Chaovalitwongse, his work with machine learning began about 15 years ago. He and fellow colleagues used electroencephalography (EEG) to analyze abnormal patterns in the brain using machine learning techniques to predict epileptic seizures. 

They also collaborated with a neurosurgeon to focus on neurophysiology by looking at neuron signals to assess how to best perform deep brain stimulation procedures in patients who suffer from tremors. They also examined where causes of attention deficit/hyperactivity disorder (ADHD) may be located in the brains of children and Alzheimer's disease in adults. 

Chaovalitwongse explained his work has shifted from focusing solely on physiology of the brain to both physiology and functional MRI data through machine learning. 

Recently, he has started using AI technology for cancer research in hopes that it could soon predict whether a patient would survive cancer treatment. 

"To detect cancer, that’s not a problem. But to predict the outcome of the treatment, I think that's harder," Chaovalitwongse said. 

He and his colleagues used AI machine learning technology to predict chances of survival or recurrence of cancer 80 to 90 percent of the time. They analyzed patterns of abnormality and extracted measurements of the shape, texture and distribution of the cancerous cells in coordination with PET imaging scans. 

Chaovalitwongse believes that deep learning and medical imaging together will be a new approach for feature generation engineering, or "how you can characterize a 3D object to describe or predict something,” such as the shape and texture of cancer cells.

The future of machine learning and medical image analysis

However, the technology will present challenges. The lack of labeled data, deficient standardization of machine learning systems between institutions, high patient costs for imaging exams and cases involving rare cancers pose challenges, according to Chaovalitwongse.  

"Not having enough labeled data, data that you know the outcome of, is problematic, especially in cancer research. Another issue is going to a different hospital with a different imaging system. There needs to be standardization between institutions. It's hard to integrate and combine data because there needs to be a standardization of the image," Chaovalitwongse said. 

But how long will it take to gather enough clinically labeled data to use machine learning in healthcare effectively? 

"I think we're getting there and I think we're recognizing that," he said. "We need more use cases and more training data. I think in five years people will start pulling data together. The National Institutes of Health (NIH) has done a great job trying to make the data public and who gets funding."