The healthcare industry is ripe for adoption of multiple aspects of artificial intelligence (AI). In a segment with an abundance of use cases to inform AI solutions, it’s easy to see how the healthcare industry can benefit from the insights provided by AI. And the stakes are high when patient outcomes can be impacted by prediction and rapid, early detection and treatment.
Intel works with several partners to develop AI solutions that address some of the healthcare industry’s most difficult challenges. One area is within 3D MRI imaging to identify and diagnose osteoarthritis. The Center for Digital Health Innovation (CDHI) at the University of California San Francisco (UCSF) has made significant strides in the development of an AI solution, backed by insights from big data and analytics, that can more effectively identify osteoarthritis on an MRI scan with assigned severity graded on a scale.
Degenerative joint disease, caused by “wear and tear” over time, leads to osteoarthritis. It occurs when the cartilage or cushion between joints breaks down and patients experience pain, stiffness and swelling. Millions of adults are diagnosed with osteoarthritis, and that number is only expected to increase. One might associate osteoarthritis with an aging population, but ACL tears are a well-known risk factor for the development of early post-traumatic osteoarthritis in a young, active population. Further complicating the matter is the challenge radiologist’s face in diagnosing osteoarthritis through Magnetic Resonance Imaging (MRI).
MRI is the most powerful, accurate, noninvasive method for diagnosing meniscal tears that lead to osteoarthritis. However, the clinical diagnosis of meniscal tears becomes more difficult and unreliable in the presence of concurrent acute ligamentous injuries of the knee and also due to the sheer complexity of the knee joint itself.
The knee is the largest and most complex joint in the human body. Vital for movement, the knee includes an interwoven structure of bone, cartridge, ligament, muscle and tissue. Within this joint is the meniscus, which is what a radiologist must have a clear view of through an MRI to determine the prospects of osteoarthritis risk, development and recovery. Because the meniscus is not clearly defined within an MRI, and there are also many borderline osteoarthritis cases, two radiologists can review the same MRI and have different opinions.
An automated system, with the help of Intel AI, is worth an investigation to see if it can assist medical professionals in decoding MRI scans.
CDHI used big data and analytics to cull hundreds of historical 3D MRI scans, and trained a deep learning model on that data to help classify the degree of meniscus degradation, which influences the likelihood of developing osteoarthritis. They included a broad spectrum of patients across various ages, genders and types of injuries, including those with meniscus tears and healthy scans without any meniscus damage present. A broad set of unbiased data was critical to the deep learning process.
In the first phase, it was necessary to use deep learning models to do multi-tissue detection (cartilage, meniscus, bone, ligaments) and segmentation. In the second phase, the CDHI team did classification of meniscus degradation with lesion detection.
CDHI examined a number of options for implementing an AI solution and chose to build and deploy an MRI classification system with BigDL: Distributed Deep Learning on CDH 5.9*, on their existing Intel® Xeon® processor based system.
BigDL is an open source library that provides rich capabilities for building solutions to help drive deeper insights with your data. BigDL can merge two paths of learning capabilities – traditional distributed analytics with optimized deep learning – for building advanced AI-based analytic systems on the Intel platform.
In this case, BigDL provided the support needed to get insights from the 3D MRI scans and made the deep learning process more accessible to the data scientists at CDHI. The programmers were able to write deep learning applications as standard Spark programs to run on top of existing clusters. This put deep learning workloads more directly in touch with the data available from thousands of 3D MRI scans, using deep learning on top of the same cluster where the data was stored.
This automated system, powered by AI, has shown great potential in helping radiologists decode MRIs to make more precise diagnoses, particularly with borderline osteoarthritis cases with only small tears.
In the future, additional factors will be added to the AI system, such as the patient’s age, gender, BMI, etc. to give further insights that will continue to inform and improve the predicted grading system.
The result of this work is an AI application that can enable faster, more accurate medical diagnoses. Deep learning algorithms, in particular convolutional networks, have rapidly become the methodology of choice for analytics in the medical imaging domain. Through open source, there’s an abundance of opportunities to share and implement data analytics and machine learning already developed for faster deployment, while the BigDL library enables the exportation of AI expertise to data scientists working across thousands of applications in hundreds of fields.
Notices and Disclaimers
Intel, the Intel Logo, and Xeon are trademarks of Intel Corporation or its subsidiaries in the U.S. and/or other countries.
*Other names and brands may be claimed as the property of others.
© Intel Corporation