Diagnostic imaging is a field where advances in machine intelligence will prove to be hugely beneficial. Early detection of anomalies in medical scans can have life-saving implications for patients.
The Intelligence Processing Unit (IPU) is efficient at accelerating medical imaging models in this way, as it is specifically designed for intensive machine learning and deep learning tasks. Beyond simply uploading DICOMs, images and CT scans, as with a CPU or GPU, with the IPU medical researchers can easily run end-to-end prediction models on large CT volumes, generate radiology reports faster and much more.
Following on from our collaboration announcement with Microsoft in November, Microsoft have released a demonstration of the Graphcore IPU working on the intracranial haemorrhage detection problem. By running these types of medical imaging models faster where timing is critical, the IPU has the potential to accelerate diagnoses and reduce time to treatment, harnessing Microsoft’s Azure ecosystem.
In the below video, Microsoft AI & Advanced Architectures Research Lead, Sujeeth Bharadwaj, demonstrates how the IPU is twice as fast and 4x more efficient than a leading GPU – while using only half the power – for ResNeXt-50 inference on a 3D CT volume.
The NDv3 series of Azure Virtual Machines backed by the IPU is available now. If you would like to accelerate your own vision models using the IPU, sign up for IPU Preview on Azure.