Computer Vision News - January 2020

3 Summary Real Time Modalities 9 of yet. The other is to be able to model confidenceandtrainingdatacompliance and uncertainties in a very reliable way, which is important for clinical decision- making. He explains: “The standard ways uncertainties are treated in the CVPR communities and computer vision communities are insufficient for what we areactuallydoing.Ageneralwaytoderive very sensible confidence metrics of deep networks would be really impressive. We currently have application-specific approaches. Most are stochastic Monte Carlo simulation approaches, not the shoot-and-forget simplistic approaches which are usually the most powerful.” Industry is another area where there is a real buzz right now. New start-ups and manufacturers are forming in ultrasound and MRI, thanks to the progress in machine learning. Bernhard says they are still trying to figure out their exact roles but have been great at facilitating relationships: “If I develop an algorithm, I would not put all my effort into getting it on to a clinical trial. A company building a product around it does that. They have a strong interest to link us academics to the clinical parts of the landscape. We are very well connected with individual clinicians but are more connected at the high level where there are problems that we can solve algorithmically . The companies are expanding and linking at a much wider scale. They are putting in huge amounts of effort, particularly engineering effort, and getting products approved by regulators piece by piece.” The big traditional imaging manufacturers - Philips, GE and Siemens - are playing a major role, but newcomers are making their mark too. He points to Butterfly Network , who do revolutionary machine learning and imaging research, and Hyperfine in New York who make bedside MRI devices. To his knowledge, none have yet completed a multi-centre large- scale clinical trial on machine learning, but he thinks we will see that in the near future.