MICCAI 2021 Daily - Tuesday

In the medical imaging field, the collection of large-scale labeled data can be very challenging without access to rare and expensive annotators with the expertise to identify abnormalities. By contrast, for natural images, such as identifying cats or dogs or the boundaries of a particular object, Amazon Mechanical Turk can provide any worker to fulfil that task. Medical data can also be noisy and imperfect due to the diversity of sensing modalities or missing entries where patients have missed a particular test. “ Data augmentation techniques are so important in medical image computing because so many problems can be solved using data-intensive methods, ” Sharon tells us. “ You can use unsupervised or self-supervised methods to leverage a large amount of unlabeled data. Data augmentation or imputation techniques can remove noise or correct mistakes in the data labels. Now, there are generative machine learning techniques like generative adversarial networks and others where you can use those models to generate synthetic data. The quality of those data can be very high, meaning they look very photo realistic. You can use the data to augment the existing training set. Techniques have shown that by using synthetic data for augmentation, performance of some medical image analysis systems is significantly improved . ” 14 DAILY MICCAI Tuesday Workshop: DALI Sharon Xiaolei Huang is an Associate Professor in the College of Information Sciences and Technology at Penn State University. She is co-organizing the Data Augmentation, Labeling, and Imperfections (DALI) workshop at MICCAI on Friday, alongside Nicholas Heller and Hien V. Nguyen. Sharon tells us more about what we can expect from the day. Data Augmentation, Labeling, and Imperfections