Computer Vision News - August 2020

Jasper Linmans is a second-year PhD student in the Department of Pathology at Radboud University Medical Center in Nijmegen, The Netherlands, where he is part of the Diagnostic Image Analysis Group (DIAG). His supervisors are Geert Litjens and Jeroen van der Laak. At this year’s MIDL conference, he presented a paper exploring a deep learning challenge in digital pathology. He speaks to us about his work. To successfully apply deep learning in medical imaging in a real-world clinical setting, predictions must be reliable . Systems should be accurate for classes seen during training and provide uncertainty estimates for abnormalities and unseen classes. However, models can behave strangely on unseen data , and this can produce very high confidence values. If you give a classification model an unseen class that was not part of its training objective, it will always predict one of the classes that it was trained for. When applying these algorithms in clinical practice, it is difficult to assess whether the output is a sensible prediction based on what it has seen during training, or if it is guessing at random . “An example of this would be if you train a model to predict one type of cancer,” Jasper explains. “When applying the algorithm, if it sees a different type of cancer, it will make a prediction based on its training, which won’t make sense for this new cancer class. It will fail silently. We don’t want models to fail silently on anomalies; we want them to fail loudly!” The aim of this research is to create algorithms that can flag such anomalies or out-of-distribution data and filter out data that the model is not able to make a valid prediction on . This data could be anything from insignificant scanning artefacts, which can be safely ignored, to clinically relevant abnormalities, such as rare pathologies. With this added information, a pathologist in a clinicalenvironment will be better able to interpret the output of the model . EfficientOut-of-DistributionDetectioninDigital Pathology Using Multi-Head Convolutional Neural Networks 34 Best of MIDL 2020 Spotlight Presentation