cool hit counter MIT Lab: training AI to diagnose diseases using just 50 images!_Intefrankly

MIT Lab: training AI to diagnose diseases using just 50 images!

TED Talk Transcript (Answer the Devil Community: Research updates, resource dry-runs, interactive Q&A)

Today's computer algorithms, are using human-like intelligence to perform, on a large scale, incredible tasks with high precision. And this computer intelligence is often referred to as AI, or "artificial intelligence." Artificial intelligence promises to have an incredible impact on our lives in the future. Today, however, we still face many challenges in detecting and diagnosing some life-threatening diseases, such as infectious diseases and cancer. Every year, thousands of patients lose their lives to liver and oral cancers.

The best way to help patients is to get early detection and diagnosis of these diseases. So, how do we detect these diseases today? Can AI help? For unfortunate people who are suspected of having these diseases, specialist physicians will first ask them to irradiate very expensive medical images such as fluoroscopy, CT, MRI, etc. Once these images are collected, another specialist physician will make a diagnosis and communicate with the patient. Obviously, this is a very resource-intensive process that requires two expert physicians and expensive medical imaging technology. This is not practical in developing countries and, indeed, in many industrialized countries.

So, can we solve this problem with artificial intelligence? Currently, if I were to use a traditional AI architecture to solve this problem, I would probably need 10,000 - I repeat, 10,000 of these very expensive medical images to be generated in the first place. Afterwards, I will go to a medical professional to analyze these images for me. Using both types of information, I can train standard deep neural networks, or deep learning networks to diagnose patients. Similar to the first step, traditional AI approaches suffer from the same problem: that is, they require a large amount of data, expert physicians, and specialized medical imaging technology.

Can we create an AI architecture that is larger, more efficient, and at the same time more valuable, to solve these important problems we face today? And that's what our team at the MIT Media Lab is working on. We have developed a variety of novel AI architectures to address some of the most important challenges we face today in medical imaging and clinical trials.

Mobile-assisted primary screening models for augmented diseases (Answer Magic Community: Research updates, resource dry runs, interactive Q&A)

The example I'm sharing today includes both of our goals. The first goal, is to reduce the number of images needed to train AI algorithms. The second, larger goal, is that we want to allow patients to use less expensive medical imaging technology. So how do we do it?

For the first goal, we chose to start with a single image as opposed to traditional AI that starts with thousands of expensive medical images. Based on this image, my team and I came up with a very clever way to extract billions of packets of information. These information packages contain color, pixel, morphology and disease features on medical images. In this way, we have converted an image into billions of training data points, and the amount of data that needs to be trained is greatly reduced.

The second goal, is to reduce the use of medical imaging techniques on patients. At first, we'll use a standard white light photo from a digital SLR camera or cell phone. And then, remember the billions of packets of information? By overlaying the information package of these medical images on this image, we have created a composite image at this point. Surprisingly, we only needed 50 - emphasis on just 50 - of these composite images to train our algorithm to be more efficient.

Machine learning for combined fluorescent biomarker and expert annotation classification using white-light images (Answer Magic Community: Research updates, resource dry runs, interactive Q&A)

To summarize our approach, unlike training AI algorithms with 10,000 expensive medical images, we use a completely new way to provide a diagnosis with just 50 high-resolution standard photos taken by a digital camera or mobile phone. More importantly, in the future, or even now, our algorithms could accept some white light photos taken by patients themselves as an alternative to expensive medical imaging technology.

I believe that we are ready to enter an era where AI is having an incredible impact on our future. I also think we should keep thinking about non-traditional AI architectures compared to traditional AI that is rich in data but difficult to apply. They are able to accept small amounts of data and solve some of the important problems we face today, especially in health care.

Dr. Pratik Shah has created a new intersection between engineering, medical imaging, machine learning and medicine. His research program at the MIT Media Lab develops scalable, low-cost diagnostic and therapeutic approaches. (Answer Magic Community: Research updates, resource dry-runs, interactive Q&A)

1、Huaweis new 40megapixel triplecamera P11 renders revealed
2、Blockchain Academy Coin Mining Chain is it a circle or a trap
3、Xiaopengs chief scientist Guo Yandong intelligent perception is the basis for improving user experience
4、Episode 100 One Day Countdown to the Rapture
5、Traffic Casualty Insurance Freight Autonomous driving may bring more improvements and changes than that

    已推荐到看一看 和朋友分享想法
    最多200字,当前共 发送