A long time ago, Artificial Intelligence (AI) gave people the same feeling as in the science fiction film. Until last March, Google’s Deep Mind developed the Go AI, AlphaGo, which defeated Go veteran Li Shishi, overnight. The reputation of AI is everywhere in the streets. People don't know AI and don't understand Go. They all mix this topic. Later, while the world of chess was still unable to let it go, it ran to play e-sports, and later, it said that it would enter medicine.
At this point, we really care about it. I don't know, it turns out that AI wants to study medicine and has thought about it for more than 30 years! Probably not very good results before, at least not "threat" to the doctor's existence, so in the eyes of our general practitioners are only small transparency. However, during this time, the activities of the big AIs have been frequent. In addition to the upgraded version of AlphaGo, the Master has gone to the round of chess to stir up a bloody hurricane, and the medical profession has frequently appeared. In the short period of three months from the end of the year to the beginning of this year, several new researches have been published in top medical journals; in particular, Nature Biomedical Engineering, which was added in Nature in January this year, has three manuals. Intelligent research is quite amazing.
JAMA: Highly sensitive, highly specific diagnosis of diabetic retinopathy
AI's "deep learning" technology consists of a series of algorithms that enable programs to perform certain behaviors by learning and optimizing themselves against a large sample data set. The algorithm used in this JAMA study is a convolutional neural network (CNN) for image recognition and classification.
The researchers used 128,175 retinal photographs as a training database to allow AI to automatically detect diabetic retinopathy and macular edema. Nearly 130,000 images were classified and graded by 54 US-certified ophthalmologists and senior ophthalmologists.
After an eight-month study period, the researchers asked AI to verify its two operating points in two new databases, one for high-specific selection and one for sensitivity. The new databases, EyePACS-1 and Messidor-2, are from two ophthalmic clinics, containing 9,933 and 4,997 retinal photographs, sorted and rated by at least 7 ophthalmologists. For the data set EyePACS-1, AI has a sensitivity of 90.3% for high specific operating point recognition and a specificity of 98.1%. For Messidor-2, the sensitivity of AI recognition is 87.0%, and the specificity is 98. .5%. Using the second operating point of AI, namely high-sensitivity operating point recognition, the sensitivity to EyePACS-1 is 97.5%, the specificity is 93.4%, and the sensitivity to Messidor-2 is 96.1%. 93.9%.
Sensitivity and specificity curves for AI in the EyePACS-1 data set. Eight dots represent 8 ophthalmologists, and the two prisms represent the high specific operating point and high sensitivity operating point of AI, respectively.
The study, published last December, shows that AI's image recognition skills have great specificity and sensitivity, comparable to human ophthalmologists. However, the researchers also said that its clinical application remains to be further evaluated. By the way, this research is also led by Google, and is carried out in cooperation with many well-known research institutions in the United States and India.
YT-T15
YT-T15
Shenzhen Sunshine Technology Co.,Ltd , https://www.szyatwin.com