Artificial Intelligence in Cancer Diagnostics, Study notes of Artificial Intelligence

Download Artificial Intelligence in Cancer Diagnostics and more Study notes Artificial Intelligence in PDF only on Docsity!

Smarter, More

Accurate, and

Less Expensive

How Artificial Intelligence

is Revolutionizing Cancer

Diagnostics

May 2020

Written by: Ethan Gauvin

Table of Contents

I. Introduction

II. Why Now?

1. The Rise of Big Data and Machine Learning

2. How Machine Learning Works

III. Improving Cancer Diagnosis

1. Improving X-ray and CT Scan Diagnosis

2. Improving Tissue Sample and Biomarker Diagnosis

IV. Economic and Social Benefits

1. Tangible Savings

2. Intangible Rewards

V. Conclusion

II. Why Now?

The concept of AI—computers capable of performing tasks that normally require human intelligence—has been around since the late 1950s. Moreover, the medical benefits of early- diagnosis have been known for centuries. So why are we just now seeing groundbreaking advancements in AI-enabled diagnostics? The answer can be primarily attributed to two emerging trends: an explosion in clinical data and the advent of machine learning.

1. The Rise of Big Data and Machine Learning

The past two decades have witnessed an extraordinary growth in the volume, variety, and velocity of data across all industries, but especially in healthcare. According to a report by the International Data Corporation, the volume of healthcare data —such as lab reports, clinical notes, wearable device data, and radiology scans—is expected to grow faster than any other sector between 2018 and 2025. [1] Doctors have gone from not having enough data on a patient to having far too much data to comprehend.

Fortunately, recent advances in AI have now made it possible to quickly make sense of large quantities of data. A subcategory of AI known as machine learning (“ML”) is particularly well-suited for big-data analysis, and is responsible for the bulk of today’s AI research, investment, and product development. Whereas traditional computing (and early forms of AI) followed a “top- down” approach requiring computer systems to adhere to a rigid set of instructions (if x input, then y output), ML takes a “bottom-up” approach allowing software to train itself on large amounts of data through trial-and-error—not unlike the development of the human brain. [2]

---

[1] Reinsel, D., Gantz, J., & Ryding, J. “The Digitization of the World.” International Data Corporation, Nov. 2018, https://www.seagate.com/files/www-content/our- story/trends/files/idc-seagate-dataage-whitepaper.pdf [2] Vigliarolo, Brandon. “Machine Learning: A Cheat Sheet.” TechRepublic, 9 Apr. 2018, techrepublic.com/article/machine-learning-the-smart-persons-guide/

Rather than code a response to countless potential outcomes, today researchers can feed data (e.g. images, soundbites, text, or games of chess) through an ML algorithm and teach it to make accurate decisions over time through millions of rounds of trial-and-error. The more data used to train an ML network, the more accurate it becomes.

For instance, if a researcher wanted to train an ML algorithm to distinguish dogs from cats, she would run it on thousands of (accurately labeled) dog and cat images. For each image, the algorithm makes a guess as to whether the creature in the photo is a dog or cat. After guessing, the algorithm is provided with the correct answer, and then records whether its guess was right or wrong. After repeating this thousands or millions of times, the algorithm will have “learned” how to identify a dog versus a cat, and will be able to do so with a high degree of accuracy.

While this process sounds straightforward, it can be successfully applied to extremely complex problems. In 2016, Google subsidiary DeepMind was able to develop an ML algorithm that beat the world champion of the notoriously complex game of Go, shocking the global AI research community.

Because there are more potential Go configurations than atoms in the known universe, it is impossible to program software to counter every conceivable move. Instead, Deepmind researchers trained their program on tens of millions of games, and over time it developed Go strategies that the best human players could not match, much less imagine. [5]

ML is particularly well-suited for medical diagnoses for the same reasons it became useful in the first place: an abundance of clinical data and cheap computing power has allowed researchers to train ML to instantly recognize patterns and pathologies that humans are incapable of detecting quickly (or at all). One can imagine how, after being trained on hundreds of thousands of x-ray images, an ML algorithm could become a quick and effective diagnostic aid. This is exactly what we are starting to see today, and ML experts believe this is only the beginning.

---

[5] “AlphaGo: The Story so Far.” Deepmind, 2020, deepmind.com/research/case- studies/alphago-the-story-so-far.

1. Improving X-ray and CT Scan Diagnosis

The majority of cancers are diagnosed through medical imaging, such as x-rays and CT scans. The challenge with this approach is that it depends on human interpretation and does not always provide a reliable diagnosis—even for the most seasoned radiologists and oncologists. Healthcare professionals routinely see a false negative diagnosis rate of 20- 30% (false positives are also common). [7] Moreover, the growing number and availability of diagnostic images is rapidly exceeding the capacity of specialists—especially in developing countries.

Even in the earliest stages of its development, ML proved it could rival the best human experts at accurately diagnosing a variety of diseases. A meta-analysis of 82 studies published between 2012 and 2018 compared the diagnostic accuracy of human professionals against ML, and found that on average ML models equaled their human counterparts in both sensitivity and specificity (sensitivity measures the proportion of positive cases accurately identified and specificity measures the proportion of negative cases accurately identified). [8]

---

[7] Bradley, Stephen, et al. “Sensitivity of Chest X-Ray for Lung Cancer: Systematic Review.” British Journal of General Practice, 1 June 2018, bjgp.org/content/68/suppl_1/bjgp18X696905. [8] Liu, Xiaoxuan, et al. “A Comparison of Deep Learning Performance against Health-Care Professionals in Detecting Diseases from Medical Imaging: a Systematic Review and Meta- Analysis.” The Lancet, 1 Oct. 2019, thelancet.com/journals/landig/article/PIIS2589- 7500(19)30123-2/fulltext.

Today, ML algorithms are beginning to outperform their human counterparts. A recent study conducted by an international team of researchers found that ML was superior to a group of 58 dermatologists at diagnosing skin cancer, catching 95% of melanomas compared to the human specialists’ 86%. [9] In 2020, Google researchers and the National Cancer Institute collaborated to see if ML could predict lung cancer by analyzing 43,000 low-dose CT scans of the lungs from 15,000 different people. For patients with only one scan available, Google’s model outperformed every radiologist who also examined the scans to assess risk of lung cancer. Google’s ML algorithm also reduced false negatives by 5% and false positives by 11%. [10]

Of course, the future of diagnostics will require both the rigorous analysis of ML and careful interpretation by human specialists. A study in 2019 found that, when reading hundreds of breast cancer screening images, radiologists’ sensitivity improved by 8% and specificity by 6% when they had an ML tool to assist them. Moreover, the time required to read and interpret the findings fell from a minute to less than thirty seconds with the help of ML. [11] Not only is ML improving the accuracy of diagnostics, but it is also improving the efficiency of human specialists, which will ultimately result in a greater number of diseases diagnosed and lives saved.

---

[9] “Man Against Machine: Artificial Intelligence Is Better than Dermatologists at Diagnosing Skin Cancer.” ESMO, 29 May 2018, esmo.org/newsroom/press-office/artificial-intelligence- skin-cancer-diagnosis. [10] Savage, Neil. “How AI Is Improving Cancer Diagnostics.” Nature News, Nature Publishing Group, 25 Mar. 2020, nature.com/articles/d41586-020-00847-2. [11] Sokol, Emily. “Deep Learning, AI Improve Accuracy of Breast Cancer Detection.” HealthITAnalytics, 1 Aug. 2019, healthitanalytics.com/news/deep-learning-ai-improve- accuracy-of-breast-cancer-detection.

IV. Economic and Social Benefits

Consider a healthy, middle-aged woman in the year 2030. She visits her primary care doctor for a routine check-up and within a couple hours is informed that a protein biomarker indicating the presence of breast cancer was detected in her saliva. While the news is unsettling, she is reassured by her doctor that tests nowadays are capable of catching cancer at the earliest stage, and that she should come in for additional screening. Her mammogram is processed by an algorithm and, although invisible to the human eye, the computer identifies a group of cells that will one day form a tumor if left untreated. The cancerous cells are confirmed by a quick biopsy, removed, and she is left with instructions to conduct a monthly breast cancer screening, which can now be done in the comfort of her own home by taking a swab of saliva and mailing it to a lab.

While it may sound like science-fiction, these types of diagnostic solutions are actually in development today. A future where routine preventative care includes non-invasive ML-enabled tests for multiple types of cancer—just as patients are routinely tested for early stages of heart disease today— could become a reality in this decade.

1. Tangible Savings

Because of their inexpensive and non-invasive nature, AI- enabled diagnostic tests will greatly expand access to cancer screening. This will result in a dramatic increase in the number of people screened each year, especially within developing countries and low-income demographics, as well as the number of cancer cases diagnosed and treated early.

The economic and social benefits of these developments will be transformational. In 2019, the cost of cancer treatment in the U.S. reached $150 billion annually, with 1.8 million new diagnoses and 600,000 deaths. In addition, it is estimated that cancer-related mortality and productivity loss cost the American economy an additional $180 billion, with the overall economic burden of cancer totaling $330 billion or 1.8% of gross domestic product. [15]

---

[15] “The Economic Burden of Cancer.” American Cancer Society, 2020, cancer.org/taking- action/economic-burden/.

Today, the five year survival rates of common types of cancer— such as breast, prostate, testicular, and thyroid cancer—are all close to a remarkable 99% when caught at the earliest stage. [18] It is not a stretch to say that, when early-diagnosis becomes the norm for the majority of cancers in the near future, millions of lives will be saved every year. While economists have attempted to assign a monetary value to an individual human life (the U.S. government puts it at $ million), the cumulative total of future lives saved is unarguably priceless. [19]

---

[18] Kandola, Aaron. “Top 7 Most Curable Cancers Based on 5-Year Relative Survival Rate.” Medical News Today, 7 Aug. 2018, medicalnewstoday.com/articles/322700. [19] Thomson-DeVeaux, Amelia. “What Should The Government Spend To Save A Life?” FiveThirtyEight, 27 Mar. 2020, fivethirtyeight.com/features/what-should-the-government- spend-to-save-a-life/.

V. Conclusion

AI and ML have only just started to revolutionize the field of cancer diagnostics. In the years ahead, they will radically democratize cancer screening, facilitate widespread early diagnosis, and save millions of lives. The subsequent economic savings and social benefits are difficult to comprehend but extraordinary in their potential.

It is important to note that these technologies will face many challenges and setbacks throughout their development and implementation. No discussion of AI and healthcare is complete without acknowledging important privacy, data governance, and ethical concerns, which carry with them intense regulatory scrutiny and political obstacles—not to mention entrenched private interests that will oppose any attempt to make healthcare more cost-effective. These concerns are outside the scope of this paper, but there are many excellent resources that cover them in detail. [20]

Lastly, while AI and ML are exciting tools with incredible potential, they must be recognized for what they are: tools. No algorithm is capable of the careful judgement, communication, creativity, and other uniquely-human literacies that medical professionals utilize on a daily basis. The purpose of these technologies is to augment physicians’ ability to solve complex problems, not to replace their jobs or decision-making.

---

[20] Rigby, Michael J. “Ethical Dimensions of Using Artificial Intelligence in Health Care.” American Medical Association, 1 Feb. 2019, journalofethics.ama-assn.org/article/ethical- dimensions-using-artificial-intelligence.