Technologies
Breast cancer detection start-up Therapixel has adopted Vast Data’s full flash storage platform to accelerate the training of its deep learning algorithms on mammography images.
AdvertisementFounded in 2013 by two Inria researchers, Olivier Clatz and Pierre Fillard, Therapixel has designed and developed technology to analyze medical images using artificial intelligence. To train its algorithms more efficiently, in 2022 it chose to rely on Vast Data’s full flash storage platformla solution Universal Storage.
In 2017, Therapixel won the DREAM Digital Mammography Challenge, an international competition in which nearly 1,200 research teams participated. After this victory, the company, which had initially designed medical imaging software for surgeons in the operating room, pivoted to specialize in the detection of breast cancer from mammogram images and the results interpretation. In order to obtain efficient algorithms, the company must train them on very large volumes of data. For example, during the tournament that launched it in the field of radiology, the teams had 640,000 anonymized images to interpret.
Speed of data access
With such volumes, storage can quickly become a limiting factor. “Previously, we relied on slower hard disk storage solutions, which did not have enough storage space for the data needed to train the algorithms that solve breast cancer,” explains Aurélien Chick. , data manager at Therapixel. “However, you cannot develop any AI solution without rapid access to data,” he points out.
It is to meet this need that the company turned to the Universal Storage platform. This has allowed it to both increase performance and double its storage capacity, with also gains in terms of scalability and ease of use. Now, Therapixel can rely on more than one million mammograms to train and test its algorithms, allowing it to detect breast cancers with a high level of reliability from the first reading and with higher accuracy. The software also provides radiologists with a real-time view of the features identified by the algorithm, to help them focus on suspicious lesions. “We can develop ideas and test them quickly, then confirm these ideas with more data to train the algorithms”, appreciates Aurélien Chick.
Article written by
Aurelie ChandezeDeputy CIO Editor
Follow the author on Linked In,
Share this article
