The Challenge: Find Abnormal Patterns in Nature
Much of AI is devoted to detecting normal and abnormal patterns in nature. An abnormal pattern is the deviation of cellular structure associated with cancer. It can be one of the most devastating patterns to find. Leukemia happens in children and adult; treatment is more effective when the disease is recognized early. New tools and techniques can improve effective leukemia detection. The combination of AI, mixed reality, spatial computing, and computer vision tools—optimized by the Intel® Distribution of OpenVINO™ toolkit—suggests an innovative approach to the problem. It adds to research for developing diagnostic tools that detect leukemia early.
The Solution: Use oneAPI Technologies to Enhance Medical Diagnoses
Magic Leap* 1, an augmented reality headset, allows users to view possible instances of leukemia in blood samples. The lightweight headset superimposes 3D computer imagery over real-world objects. In this case, it used images from a blood sample. AI training data used images labeled by expert oncologists as likely instances of Acute Lymphoblastic Leukemia (ALL: sometimes called Acute Lymphocytic Leukemia). The presence of probable ALL lymphoblasts determines a positive classification.
ALL is a blood cancer characterized by fast growth. When this disease affected a family member, Adam Milton-Barker devised a potential method for precise detection of the cellular patterns typical of ALL. His medical diagnostic method takes advantage of OpenVINO toolkit components, the Intel® Movidius™ Neural Compute Stick 2 (Intel® NCS2), and several other Intel® technologies and tools.
"I want to use technology to make a difference [in] the medical industry," Adam said, "and build a team that will help conquer early detection of leukemia so that families may not have to go through what my family went through."
Through the organization he founded, Peter Moss Leukemia MedTech Research, Adam hopes to forge relationships with hospitals. His goal is to get large volumes of data to further validate the classifiers and develop a real-world implementation of the diagnostic tool. The organization was founded as a tribute to his grandfather who had a terminal diagnosis one month after an all-clear blood test. He died one year and one day later.
"The biggest issue," Adam said, "is not so much technical but rather that data for Acute Myeloid Leukemia is hard to find. We have mostly focused on Acute Lymphoblastic Leukemia due to the lack of open source data for AML."
"The other major challenge," Adam continued, "is the fact that early detection of leukemia is a global problem that has no solutions. This is why our first research project started, and our aim is to use as many types of technology, frameworks, programming languages, and so on, as we can to try to find a solution."
Adam has a diverse background. He started as a website developer that promoted DJs and ran a forum that connected UK DJs to Spanish DJs. Once his website career began to take off, Adam left DJing to work as full-time website developer.
"I got into the Internet of Things [IoT] around 2014 [and] was awarded Intel® Galileo by Microsoft Windows* Developer Program for IoT for a project idea," Adam said. "My early days of IoT and artificial intelligence experience were spent working on automation combining natural language understanding with IoT and websites and online systems."
His career accelerated after becoming a semifinalist in the IBM* Global Mobile Innovators Tournament; a first-phase winner in the Arduino* and Microsoft* World Maker Challenge; and community teaching assistant for MIT's computer science courses.
"I was awarded Intel Experts Award at IoT World Congress Hackathon Barcelona, which led to me be being invited to join Intel® Software Innovator," Adam said. "My overall focus has been combining multiple types of technologies—IoT, AI, virtual reality, mixed reality, blockchain, online systems, to create intelligent/smart networks. I believe automation can improve business and quality of life."
Methodology and Approach
Adam developed the ALL Detection System for Magic Leap 1. It applies deep learning technology—embedded on an Intel NCS2—using a headset from Magic Leap as a mixed-reality view screen and supplemented by several supporting technologies. Initially, the technology was a proof-of-concept for using mixed reality and computer vision to enhance diagnostic effectiveness. The technology can now manage a greater volume of leukemia data, which moves the project to the next level of development.
The Intel NCS2—a compact, fanless device—delivers the solution's deep learning capabilities. Intel NCS2 provides intelligence at the network edge in a power-efficient package that's designed for exceptional performance per watt. The project was shaped using components from the OpenVINO toolkit, which provided built-in computer vision capabilities and streamlined the development process.
Combining deep learning technologies from Intel with the Magic Leap 1 headset created a unique way to conduct diagnostics tests with computer vision and mixed-reality elements. Providing pretrained models on the Open Model Zoo and important support for Raspberry Pi* hardware, the OpenVINO toolkit simplified the process of developing the key parts of the diagnostic tool.
The ALL oneAPI Classifier was the classification model. Training consisted of peripheral blood sample images from a dataset using the ALL Image Database for Image Processing, compiled by Fabio Scotti, a professor at the University of Milan.
Adam developed an interface with C# and a Unity* 3D game engine, and deployed it to the Magic Leap 1 hardware. He developed code with the Intel® Distribution for Python* programming language, and trained the classifier with Intel® Extension for TensorFlow*. Once the model was fully trained and frozen, he used the OpenVINO toolkit. After it was converted into a deep learning network intermediate representation, the solution was deployed to Raspberry Pi version 4 hardware. NSC 2 delivered inference at the edge.
Although Adam has successfully created the proof-of-concept version of this diagnostic device, more widespread acceptance for the technology—based on a greater volume of data—is needed to improve the device.
Adam said, "The actual detection of AML is an unsolved problem, and this will be an ongoing project—it is the reason our first volunteer project started and why our nonprofit was formed. We need vast amounts of data to really determine if a particular model is accurate enough to be used in real-world applications. We need to find ways AI can be used to solve early detection. These are the two main challenges that remain unsolved."
"The project is already open source," Adam said. "The mixed reality side is pretty much ready. However, it will adapt as time goes on. We hope to attract more researchers to help us work on the classifier side and form partnerships with hospitals that can allow us access to the data we need. We hope to run a pilot project in the future when we are confident that we have a good enough model."
Full code of the classifier portion of the project and the mixed-reality portion of the solution is available on GitHub*.
"The next version of the classifier is in development," Adam said, "and will use Intel® AI Analytics Toolkit, Intel® Optimization for TensorFlow*, and TensorRT*, and will run on an NVIDIA* AGX Xavier board, hosting an API endpoint that the mixed reality application will communicate with. This version will be a non-HIAS version, which will mean that users do not have to install the HIAS server to use it. This version will also support Windows."
Adam's organization actively works on several projects. Many are devoted to using the latest technology to improve medical diagnostic capabilities.
Resources and Recommendations
For the latest information on leukemia diagnostic techniques, see the Peter Moss Leukemia MedTech Research.
Explore a database of open datasets and code.
Learn more about Adam's project at Intel® DevMesh.
To learn about possibilities that oneAPI gives developers, see oneAPI.io.
Get an overview of the Intel Distribution of OpenVINO toolkit.