Author: Adam Benzion
It’s no secret that artificial intelligence (AI) — or, as my nerdier friends call it, “the outcome of really good machine learning (ML)” — is changing the way we live; the way we shop, eat, sleep is all monitored and enhanced by AI, and now the medical industry is tapping into the capabilities of AI. Entrepreneurs, hackers, and engineers have discovered new ways to infuse off-the-shelf hardware with ML tools to deliver highly specialized and expensive medical diagnostics in a field that is known for costly equipment and high barriers to entry. Welcome to the new era of AI-powered medical care.
Until recently, standard skin cancer lesion screening methods such as radiological imaging (x-rays) could miss the early signs of cancer in approximately 25% of cases, according to research published in the British Journal of General Practice. And in places where there is no access to expensive medical equipment, patients and doctors rely on experience and the naked human eye to perform early-stage cancer diagnosis. Not only does this present a non-standard level of diagnostics, but it also draws attention to the soft belly of medical care — only those who can afford quality treatment will get it. Either way, we are finding that there’s a need for other detection modalities that we’ve never tried before, breaking from traditional methods, and offering the same, top-quality early detection and diagnostics of cancer to everyone equally.
Medical care meets computer science
Meet Mohammed Zubair, associate professor at the Department of Electrical Engineering and Consultant at the Center for Artificial Intelligence at King Khalid University in Saudi Arabia. Zubair is accustomed to operating at the intersection of medical care and computer science and has been on a quest to solve one of the most critical problems in one particularly painful form of cancer affecting the oral cavity. Through his research and personal observations, he found that the biggest factor affecting oral cavity cancer patients was the time it took between initial detection and accurate diagnosis. When oral cavity cancer patients did not have fast access to experts, and too much time passed between the formation of cancerous cells and their detection, the odds of successful treatment and survivability decreased dramatically.
Zubair started his journey by mapping out the root of the problem, including long wait times to see specialists (up to months), lack of access to care in remote locations, and human error factors. All of these delays in detecting malignant tongue lesions could result in the avoidable death of a patient. Zubair decided that he would apply his expertise in this form of cancer to design a fast and inexpensive pre-screening solution that could be built faster and cheaper, and used everywhere. He wanted to develop a system that would reduce the time it takes to train physicians and even dental specialists, regardless of location and socioeconomic disparity.
To begin, he created a data warehouse with a clinically annotated tongue lesion dataset, and trained a detection model based on the Edge Impulse TinyML engine and Google’s TensorFlow Lite framework. He then ran his trained software on his iPhone, as well as a microcontroller integrated camera, using off-the-shelf computer vision to screen patients. The systems showed both benign and precancerous conditions, alongside traditional diagnostics methods. What he discovered was nothing short of astonishing. His detection accuracy (percentage of correct prediction) was close to 90%, with detection inference time at around seven milliseconds.
Zubair is validating with his research what many in his field are also saying: AI can greatly improve the accuracy of cancer detection in images, and open up new ways of developing and deploying cancer screening with a new set of bionic eyes that can see and detect what a human can’t. With enough data to train neural networks, cancer detection is being democratized, achieving impressive results.
Developments like this impact more people than you can imagine. According to the World Health Organization, more than 650,000 cases of tongue cancer are reported each year. It is prevalent in individuals mostly from developing countries due to lack of awareness and limited access to clinical diagnosis and dental specialists, predominantly in South Central Asia. This painful and debilitating disease normally manifests on the lips, gums, tongue and inner lining of the cheeks, roof and floor of the mouth. A patient suffering from this form of cancer has difficulties with eating, speaking, the appearance of lumps in the oral cavity, physical marks on the face due to surgical procedures and treatment, and severe pain.
“Automating the initial screening process for oral cancer patients using AI to detect precancerous tongue lesions can prove to be an effective and inexpensive technique that would allow patients to be triaged accordingly to receive appropriate clinical management,” Zubair said.
Rendering the human eye obsolete
Zubair’s experience is a bright spot in a new, blazing trail of experiments that are changing the way we look at medical care and cancer detection. In 2018, a team from Germany, the United States, and France taught an AI system to distinguish dangerous skin lesions from benign ones, across 100,000 image datasets, using a convolutional neural network (CNN), while comparing their hypothesis against 58 dermatologists from 17 countries. All were shown the same photos of malignant melanomas and benign moles. Once again the results scored a hands-down win for software. The AI-powered system was able to accurately detect cancer in 95% of images of cancerous moles and benign spots, whereas a team of 58 dermatologists was accurate 87% of the time.
With the right hardware, ML models, and datasets, medical-grade AI will detect other skin conditions, from pigmented lesions to rashes. However, pigmentation, lesions, and rashes can often look identical, whether they are cancer, toxin, or allergic reaction, so it’s the nuances that will make it effective, as well as the tiniest details within the dataset. Without a doubt, cancer diagnostics and detection will be best performed by AI in the coming years, even if we’re not there yet. Even untrained developers are pushing the limits on what we can do to beat this awful disease, which is always a leading indicator of things to come, see the Ultra96 Skin Cancer AI detection system shared on Hackster last year.
According to an article published by the radiological Society of North America, advancements in medical-purpose AI and deep learning are also applied to many other fields of cancer detection. In particular, applying deep learning to digital breast tomosynthesis (DBT: an advanced method for cancer detection in which an X-ray arm sweeps over the breast, taking multiple images in seconds) improves cancer detection. It reduces false-positive recalls compared to screening with digital mammography (DM) alone. With a new deep learning system mining vast amounts of data to find subtle patterns beyond human recognition, researchers tested its performance on real patients with and without cancer for comparison, detection sensitivity increased from 77% to 85% while cutting detection time by half.
We’re experiencing a healthcare revolution, powered by purposeful AI, from Google labs who recently shared an AI system that can detect the presence of breast cancer more accurately than doctors, to newly formed organizations such as CancerAI, which is committed to helping organizations to develop AI technology to dramatically improve cancer prevention, detection, and treatment.
AI-Powered Healthcare Is Our Greatest Ally
Cancer is not one disease; it’s a group of over 100 diseases involving abnormal cell growth and is the second leading cause of death around the world, taking 10 million lives annually. Not surprisingly, seven out of every ten deaths from the disease happen in low-or middle-income countries, and generally with people who cannot afford quality care. AI truly changes the way we can build, reach, detect, and possibly treat cancer patients in greater numbers at a lesser cost. AI, done right, can become humanity’s best friend after all.