MICCAI 2021 Daily - Tuesday

In recent years, vast improvements in imaging technology, as well as the variety of different methods available for scanning patients, have placed a heavy burden on radiologists , who must sift through high volumes of medical data to detect abnormalities. This work aims to use outlier detection to bridge the gap between the increasing capabilities of the imaging technology and this time-consuming data processing task. “ Outlier detection is challenging because it’s not trained to do a specific task like detecting one known class of disease, so it’s hard to validate for clinical use, ” Jeremy explains. “ Even methods that do detect specific things, like breast cancer screening, require extensive validation with a dedicated team that specifically designs the test cases before they can enter clinical practice. There are so many issues like domain generalization that have to be overcome before it can be really useful. Many of the same techniques would have to be applied for outlier detection as well. ” The most standard way to do this is by using autoencoders to form a reconstruction loss. The reconstruction can be compared with the test image input to see which parts of the image the network was unable to accurately reconstruct. This works best if there are very obvious differences in the images. However, the advantage of using a neural network is that it can pick out small subtle features and make semantic distinctions, even if the two images are not that different. This work aims to use the network to look specifically for those semantic differences and find more subtle deviations and irregularities. 18 DAILY MICCAI Tuesday Poster Presentation Detecting Outliers with Poisson Image Interpolation Jeremy Tan is a PhD student at Imperial College London under the supervision of Bernhard Kainz. His paper proposes a self- supervised method for outlier detection in medical imaging. He speaks to us ahead of his poster session tomorrow (Wednesday).