Deep Learning for Cancer Treatment

Published: February 18th, 2019

Category: Featured, Spring 2019

By: Rachel Lynch

Patients suffering from cancer can often feel like they have little to no options when it comes to treatment. Due to the nature of cancer, over time the illness evolves and becomes immune to popular treatment options. But what if we could predict how cancer will affect a patient, or how a patient will react to a specific type of treatment? Researchers are finding new ways for software to aid the medical field thanks to advancements in new technology, mainly artificial intelligence.

Researchers are finding new ways for software to aid the medical field thanks to advancements in new technology, mainly artificial intelligence.

How does AI work?

Artificial intelligence is common in everyday life and seamlessly integrated in a variety of applications, but the majority of us might not know how it really works. In summary, AI allows machines to learn from experience and operates through deep learning and complex data sets. Software engineers add data to the system, the system learns this data and then is able to make inferences based on the information it is presented. It also runs on natural language processing, so it can communicate in layman’s terms, rather than in the code it runs on. In using this software to its full advantage, today’s technology is able to process large amounts of data, accomplish tasks, answer questions on the spot, and pick out patterns in data sets seamlessly. The more data an AI computer is presented, the more intelligent the system becomes.

AI and Cancer Detection

One of the most important things AI has the potential to be used for in the medical field is cancer detection. While this is something that is still in the testing phases, the results can be groundbreaking for oncologists and cancer patients. Multiple researching groups in both Europe and the United States have tested this technology and found that on average AI can be more accurate in detecting cancer than oncologists. In one of the studies, researchers discovered AI software had the capability to detect cancerous moles vs. benign blotches when shown images of the two. While doctors were able to detect cancer 87% of the time when looking at the images, AI software was able to identify cancer 95% of the time.

With this technology on the horizon, many doctors are intrigued by its potential to help cancer patients with a poor prognosis on their hands. For a cancer diagnosis with a low survival rate or short life expectancy, such as liver cancer with 83.9% of patients losing their fight within 5 years after diagnosis, or mesothelioma, a rare cancer of the lungs with around 40% of patients living past one year to 9% living longer than five years, artificial intelligence software could greatly increase early detection and appropriate treatment plans. If oncologists were able to detect cancer at an earlier rate, there would be a higher likelihood for successfully treating cancerous cells and increasing life expectancy rates for notoriously quick moving illnesses.

AI and DNA

The DNA of a cancer cell holds the answers to solving the mystery of how to most effectively treat this disease. For some types of cancer to form, there is a DNA mutation that happens in the code of a cell. This tells the cell to continue to multiply and grow causing cancerous tissue and tumors. By looking at a cancer cell’s DNA using AI software, doctors can find the code that has been changed and use that information to predict how an individual cancer will act.

Inception-v3 a software made by Google, is able to do just that. Developers from the New York University’s School of Medicine, who trained Inception-v3 added images and information from The Cancer Genome Atlas. They found that when the technology was programmed to detect cell mutations, it was able to detect six (STK11, EGFR, FAT1, SETBP1, KRAS, and TP53) from pathology images, with only a 0.733 to 0.856 AUC when measured within a control group.

This information would allow doctors to digitally input a cell, detect whether it is cancerous or not, and tell what mutation caused this cancer. Then this technology would take that data and compare it to a database of other cancers with similar mutations to get recommendations for what types of treatment worked best. This could also help doctors predict what a realistic prognosis would look like. While doctors are able to provide much of this information on their own through tests and screenings, AI software could help make this happen quicker and improve the accuracy of cancer predictions.

Kickback on AI

AI can make great improvements to cancer and the medical field as a whole, but of course, the system is not at a level of perfection just yet. IBM tried to conduct their own version of AI software for medical use and claimed that it was able to identify a patient’s ailment, based on symptoms provided. When tested, it was unable to identify a heart attack, torn aorta, or angina when given chest pains as a symptom. Instead, it came to the conclusion that it was an extremely rare disease rather than providing the more common answer. Come to find out, IBM hadn’t used guidelines by medical associations, which could explain the machine’s incorrect outcomes.

This is something that developers are now aware of. They realize that nationally recognized medical information needs to be the power behind these machines in order to make them worthy of worldwide use one day. Furthermore, doctors need to use their common sense and expertise when working with this software, just in case there is a glitch or lag in the system.

Every day we are one step closer to curing cancer through technology. As more information and data is compiled on cancer patients and their experiences with treatment, the better AI technologies can operate for the use of future detection, prognoses, and treatments. However, medical researchers need to adapt, and understand the inner workings of AI software before cancer patients can benefit from all that technological advancement has to offer.

Edited by: Meagan Sullivan