A group of scientists from Stanford University have developed a neural network which based on several x-ray can diagnose pneumonia, and it makes it no worse than practitioners-radiologists.
To create an algorithm, scientists have created a 121-level neural network and trained it on 112 120 front fluorography of the chest, obtained from 30 805 patients. Each image was indexed in accordance with existing lung disease. In the learning process images scanned, converted to digital form, compressed to the size 224 × 224 and “feed” the neural network. Then randomly from the entire database was selected 80% of the shots and clearly setting algorithm. The remaining 20% was left to validate the system operation and debug.
In the next phase of trials, researchers from Stanford took a new 420 pictures, which gave the conclusion practitioners. The doctors had a fairly decent experience from 4 to 28 years. Neither the radiologists nor the neural network did not have access to the medical records of patients. Had only pictures. It was found that the neural network is not much inferior to the professionals with years of experience.
Curve of errors committed by physicians and by the neural network. Red crosses marked by medical errors, green error of the neural network
The researchers note that in the course of the test was used only frontal images while in medical practice are also studied side shots and medical history that has influence on the diagnosis. However, such programs help diagnose diseases in remote regions with a lack of qualified personnel.
The neural network puts diagnoses better than doctors
Vladimir Kuznetsov