Computer Vision News - August 2020

is rendered completely useless. This problem could be qualitatively observed through results from the recent FastMRI challenge at NeurIPS 2019 , where some of the top performing models were shown to have failed in reconstructing some relatively small abnormalities, such as meniscal tear and subchondral osteophyte. The team worked from the basis that clinicians need images they can rely on . They would not accept that features could not be reconstructed just because they were washed out during the under-sampling procedure. Their work addresses this false-negative problem by building on existing methods to accelerate MRI reconstruction and using dataset knowledge from training these deep learning models to recover this lost information even after under- sampling . How do they do it? First-author Victor explains: “We use adversarial attacks and robustness training . We frame these small abnormalities as adversarial attacks that the deep learning methods are missing. It’s like a feature that is constantly trying to poke the algorithm to make it fail to reconstruct. By constructing it in this way, we can find out all the ways that the algorithm failed, and then patch them through additional training . That’s basically what we did, and it turned out to be useful in terms of defending against these types of synthetic attacks. Also, in real-world reconstruction we observed some improvement. This is just the first step towards addressing not only the image quality but also the diagnostic features to reconstruct.” Fran cesco adds: “Another reason this method works so well is that adversarial attacks and robustness are areas of deep learning with theoretical guarantees on what you can do and performance, such as how much attack you can defend against with a particular deep learning system in the way that you train it or perturb the image. That’s why if we apply this kind of robustness framework to MRI accelerated reconstruction, we can actually provide some guarantees in the future that were not provided in the past. ” The team have observed that the acceleration factor can go way beyond what has come to be accepted Compressed sensing has been around for almost 15 years, but with such a low acceleration factor, its use in clinics is not widespread. A 2x or 3x increase is not enough of a pay-off for the loss in image quality. With these new Victor and Francesco 13 Best of MIDL 2020