Deep learning assists technicians analyse medical images
A trial using deep learning algorithms has shown that artificial intelligence has the potential to assist technicians and detect human errors in medical image handling.
System-on-chip manufacturer Socionext and Japanese AI software company Soinn presented results from the project at Medtec Japan, held in Tokyo from 19-21 April.
In the trial, Socionext extracted and delivered biometric data to Soinn’s Artificial Brain. Soinn learned to read subcutaneous fat thickness from abdominal ultrasound images. The estimations by Soinn were then compared with the reading results by ultrasound technicians.
Soinn’s Artificial Brain can accurately read fat tissue thickness from 80 per cent of the data within 5 per cent margin of error. There were noticeable differences between the readings by human and by Soinn for some of the images.
After reviewing these data, it was confirmed that human error, including numerical input, was a common occurrence from data reading by human. Based on the findings, the companies believe that AI has the potential to be used for assisting technicians in reading images and detecting human errors in medical image handling.
Machine deep leaning, which is attracting attention from fields including medical imaging to driverless cars, is thought to require hundreds of thousands of images in order to learn from reading the images. In contrast, Soinn needed only about 700 images.
