Medicine

How artificial intelligence learned to diagnose cancer | by Botkin.AI | Botkin.AI | Oct, 2021

A story about how AI is changing the healthcare sector and frequently asked questions.

In 2020, among all cancer diseases, the most prominent cause of death was lung cancer 18% of all cancer deaths, to be precise. The probability of dying to this disease directly correlates to the stage of the cancer when it was discovered. For instance, if the disease was discovered during the first stage, the patient’s chance of surviving the next 5 years is 92%; and is the disease was discovered during the fourth stage, the chance plummets all the way down to under 15%. In short, the patient’s survival depends primarily upon when the disease was discovered, and whether it was discovered at all.

The latter point has proven to be problematic: studies* have shown that radiologists miss the onset of cancer on the CT-scans in more than 70% of cases.

It’s not that we’re dissatisfied the healthcare system — quite the opposite. However, we would like to point out that, in spite of rapid development of diagnostic tools, we are still unable to detect lung cancer with 100% consistency. The reasons for this are varied — high workload or shortage of qualified personnel; but the main reason is the fact that the early stages of cancer look like small lumps (sometimes as small as a few millimeters) which are exceedingly hard to detect. A late diagnosis significantly increases the treatment costs. For example, treating the stage 4 cancer costs 30 times more than stage 1 or stage 2 cancer.

Alright, we have stated the problem — now we need to find a solution. And we did. Back in 2017, our team saw the solution in the artificial intelligence.

That is how Botkin.AI was born — in 2017, we focused our efforts on medical image analysis. We began with thorax CT-scans. As of now, Botkin.AI can analyze not only the thorax, but also the head and mammography results. And to this day, we continue to expand our repertoire.

What can Botkin.AI do?

We created a platform, capable of processing medical images, detect and describe the signs of cancer. The analysis is performed via artificial intelligence, which can detect the pathological signs in the early stages. Botkin.AI is capable of both a retrospective analysis (re-processing of an archived image) and a prospective analysis (also known as on-the-spot analysis: a fresh image is sent to the cloud, where it is processed, and the results are sent to the operator)

The platform can be integrated with the image archives of one or several hospitals, and works exclusively with the images, without fetching the patient’s personal data. The image is processed in under 6 minutes, and is sent back, with all the pathological signs highlighted and described.

What exactly happens to the image being analyzed?

From the archive, the image is sent into the cloud, where it is processed via artificial intelligence. The algorithms determine any abnormal spots on the image. The whole process takes from 1 to 6 minutes. Afterwards, the doctor may log in into our cloud viewer interface and see the results.

The first page contains the list of all scans, each of them marked with a green or a red circle, depending on whether any abnormalities were detected. The rest of the pages contain individual scans, with any pathological signs highlighted. Each page also contains a side panel, where you can see the important numbers, such as the tumor size and density.

So, let me get this straight — my scans can be viewed by any IT-people in the company?

Not exactly. All images are loaded into Botkin.AI from the relevant medical archive — that much is true. However, from that point on, they are processed without any human input. Even if an employee decides to intervene, they would not be able to determine, whose scan they are looking at. All images are depersonalized — patient’s name and surname are replaced with a personal ID. And that is how we ensure that your personal data is safe.

How do we train our mathematical model?

We start by choosing a disease. This choice largely depends on how prevalent a disease is among the population; how difficult it is to detect and what is the chance of a false negative during the diagnostic procedure. We discuss with our medical experts what exactly would we need to measure. During the preparation step, we create a markup plan and decide, what criteria would we use to diagnose a disease and what would our markup tools look like. During that time, our data-scientists look up open data sets — international data sources with the data already marked up — and train the first AI generation using those data sets. We also collect data from a select number of hospitals. We take utmost care to ensure that the markup of each image matches our markup plan and that each image is reviewed two, sometimes even three times. Only then the image is used for AI training.

As of now, our datasets contain more than a million of processed medical images

What if the AI makes a mistake?

Such a thing is indeed possible. Unlike a human doctor, our platform does not process the entirety of information needed to correctly diagnose a patient, such as patient’s medical history. Our model only processes the medical image, which is why it cannot be used as a standalone diagnostics tool, and which is why we need input from certified medical personnel. On the flipside, AI can detect abnormalities that the human eye is likely to miss.

According to our data, the usage of Botkin.AI increases the chances of detecting malignant tumors to 50%.

Can you give us an actual case involving your platform?

During the Pandemic, hundreds of thousands of Russian citizens made CT scans of their thorax. The workload of medical personnel was astonishing; the doctors had to determine the rates of infection and the right treatment in record times. With such an influx of patients, it is very easy to miss signs of lung cancer, especially in the early stages. That is why we decided to organize a retrospective analysis of thorax CT scans (via the means of artificial intelligence). This project was launched in Nizhniy Novgorod and was concluded in the August of this year. We managed to collect 9,7 thousand scans; our platform has detected signs of cancer on 284 of them, with 113 patients among them never having sought oncological treatment. Those patients were then directed for further diagnosis.

Around 2020, our portfolio already had successful pilot and commercial integration projects into medical establishments, both governmental and private, in Moscow, Saint Petersburg, Chelyabinsk, YNAO and Murmansk. We also managed to launch international projects with our colleagues in Egypt, Uzbekistan and Brazil

Will the AI take away the doctors’ jobs?

Absolutely not. We strive not to replace doctors, but to provide assistance. Russia suffers from a severe lack of radiologists, especially within the regions. Oftentimes, a young radiologist has no one to turn to for advice on how to correctly diagnose a patient — sometimes the results have to be sent into a nearby city for a review! If we could integrate an image analysis platform into those regions, the doctors would be able to rely on the platform, without the need to wait for help from their colleagues. This will not only reduce their workload, but also increase the accuracy of the diagnosis.

Another promising venue is the analysis of mammography results. To correctly diagnose a patient, the results of mammography must usually be reviewed twice, and Botkin.AI may be used as a second reviewer. However, to make this development a reality, some legal changes would need to happen, because as of now, according to the Decree of Ministry of Healthcare, both reviews must be made by human doctors

*According to the research made within the framework of International Early Lung Cancer Action Program (I-ELCAP)


Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button