A collaboration between Intel Labs and Perelman School of Medicine at the University of Pennsylvania (Penn Medicine) is advancing diagnostic medicine through a specific machine learning architecture called federated learning. Unlike other machine learning models that send data to a centralized location, a federated workflow moves a model to the data for training. From there, it recombines to build a global model.
This machine learning model aims to collect valuable data into benign and malignant tumors across a number of hospitals while securing patient information.
Intel’s Focus on Machine Learning
It’s no secret that AI requires plenty of processing power to operate effectively. Federation architecture requires a few models to work simultaneously and those algorithms might be analyzing millions of data points at scale—putting a strain on computing power.
Federated learning architecture. 1 is local model sharing. 2 is global model sharing updates. Image used courtesy of Intel
Intel has developed its machine learning technology over the past few years, equipping it for this biomedical endeavor. Just two years ago, Intel shared that over 50% of enterprise AI customers leaned on their Xeon processors. We might expect a similar implementation here at the collaborator level. Intel also bought Habana Labs last December—in an effort to bolster their AI offerings. This brain tumor project could be the fruit of that acquisition.
How Intel’s Hardware and Software Will Drive the Project
The press release reports that on-site data centers for the project will use Intel hardware. Another standout technology is the company’s Movidius Myriad X visual processing unit (VPU), which supports advanced deep learning and concurrent workloads. Intel also offers dedicated AI processors and Field Programmable Gate Arrays (FPGAs) to support machine learning.
Movidius Myriad X chip. Image used courtesy of Intel
Intel’s nGraph, Data Analytics Acceleration Library, and PlaidML (among others) will support performance on the software side.
The tumor AI will soon train from a comprehensive database, via the International Brain Tumor Segmentation (BraTS) challenge. Medical professionals can thus access veiled records from which research insights can be gleaned.
Data Capture and Privacy: A Balancing Act
AI and privacy are often said to have an antagonistic relationship. A Fortune article on privacy and AI explains that heightened data capture can open more opportunities for security breaches—and yet, more data can drive greater success in machine learning algorithms. Intel and Penn Medicine hopes to address this challenge by “99% accuracy” compared to their non-private counterparts. This assurance of reliability may accomplish a few things:
- Allows researchers to collaborate with fewer restrictions
- Keeps patient records private without potential HIPAA violations
- Ensures that future patients will receive accurate care without compromises
This federated learning architecture ensures that data isn’t processed at a site vulnerable to breaches or attacks. The model travels to the data, relaying updates from localized models to a globalized model. This centralized model evolves, becoming a hybrid of its localized counterparts.
Intel and Penn Medicine acknowledge that hospitals house secure databases that are constantly changing. Researchers can share this data between research locations without the risk of interception.
Intel recently released a video discussing their role in AI development, explicitly stating cancer research as one of their goals. Screenshot used courtesy of Intel
As each model is exposed to more data, any algorithmic changes are transferred without compromising confidential patient information.
Hospitals Participating in AI-Driven Tumor Research
Penn Medicine is spearheading the effort alongside a coalition of more than 29 global research and healthcare facilities. Some noteworthy institutions kicking off phase one include:
- Hospital of the University of Pennsylvania
- Washington University in St. Louis
- The University of Pittsburgh Medical Center
- Vanderbilt University
- Queen’s University
- Technical University of Munich
- University of Bern
- Kings College London
- Tata Memorial Hospital
That National Cancer Institute has awarded a $1.2 million grant to Dr. Spyridon Bakas, head investigator, to accelerate the development of artificial intelligence models. These models will intake both patient and research data from numerous sources, learning continuously, while protecting private information.
The Benefits of AI and Rapid Detection
Effective brain tumor treatment hinges on early detection. Malignant brain cancers can spread throughout the central nervous system, especially when left unchecked. This new research collaboration will hopefully lead to improved outcomes. According to the American Society of Clinical Oncology, the five-year survival rate for cancerous brain tumors is approximately 36%—and this percentage decreases with age. Artificial intelligence algorithms can help doctors detect these covert growths before they cause preventable harm.
Mitigating Future Health Concerns
The American Brain Tumor Association estimates that nearly 80,000 brain tumors will be diagnosed this year. Improving outcomes for the vast majority of patients is the ultimate goal. This new system from Intel and Penn Medicine holds promise for honoring both HIPAA laws and inputting valuable data, yielding faster diagnoses.
Credit: Google News
