TAIPEI, Taiwan - Following the digitization of histopathological slides, the slides have extremely high resolutions with each digital microslide reaching billions or even tens of billions of pixels in resolution and the file sizes exceeding 10GB. This is not only a huge challenge in terms of storage, but also relatively time-consuming to train AI models.
AetherAI and National Applied Research Laboratories (NARLabs) National Center for High-performance Computing (NCHC) are collaborating using Taiwan Computing Cloud (TWCC), which is equipped with medical image optimization architecture, in order to realize technology which employs unified memory and graph optimization. Their challenge is to directly develop an AI model entirely through the use of digital microslides of pathological images. Each project can save half a year worth of labeling time for medical professionals, and digital pathology AI technology is ushering in enormous innovations.
AetherAI has utilized unified memory and graph optimization technologies in 2019 to tackle the challenge of training deep neural networks entirely through the use of microslides to replace the current method which relies on the labor of medical personnel for labeling and cuts the images into multiple blocks. This saves professional pathologists between a few months and six months of time spent on labeling.
Unlike the usual cloud for commercial use, TWCC makes use of a high-speed computing architecture design of which each node has a bandwidth of 100G. This facilitates the processing of smooth exchanges of full-resolution image data and makes it possible to carry out AI training entirely through digital microslide images made up of billions of pixels.
Through the increased speeds of aetherAI’s software and TWCC, the calculation speed is increased by 275 times with nearly one thousand hours of labeling time saved per project for professional physicians. In addition, through 400% increases in the ultra-high resolutions, the most complete zero-cutting method is utilized to carry out comprehensive reviews, dramatically improving the efficiency and quality of identification.
(TR/ Phil Sweeney)
