Computer Vision News - May 2019

14 Computer Vision News Challenges in Airways Segmentation Every month, Computer Vision News reviews a successful project. Our main purpose is to show how diverse image processing techniques contribute to solving technical challenges and real world constraints. This month we review challenges and solutions in a Medical Imaging project by RSIP Vision : Airways Segmentation with Deep Learning . Project Airways Segmentation is a very useful development, which helps physicians in a multitude of tasks: diagnosis, operation planning, assessment of treatment results and more. It also presents several challenges that need to be solved, before the project is able to provide its benefit. Some of this challenges are similar those found during other body parts segmentation. One of these is the wide variations of the anatomy between one patient and another: even though they are composed of the same parts, the details of these parts differ in many ways and this difference is mirrored in the CT scans. CT scans come in very diverse resolutions and quality, depending on the machine involved and its producer. The level of noise might change from one vendor to another and the algorithm needs to provide an accurate segmentation also when dealing with many of them. Most often, CT scans are done on unhealthy people, whose anatomy is different. For instance, in a patient who has had surgery before, a metallic implant might cause artifacts in the CT. Finally, the running time is a challenge too: the analysis of the medical scan needs to be fast. Some times it can run in a few seconds and in other cases several minutes are needed. RSIP Vision has a very long experience in solving these challenges. We have built a unique data augmentation system, able to generate additional CT scans, which look like real even when the anatomy has been deeply modified. Accurate annotation requires a huge effort. Even though we use partially automated tools, it takes radiology experts years of man work to annotate and validate a sufficient quantity of data. RSIP Vision has a long experience in solving these and other challenges. We always use statistical solutions, based on machine learning and deep learning, rather than heuristic methods: the latter do not offer a proper solution when a the dataset includes very diverse cases. Instead we train and validate our solutions on a varied and balanced dataset, including thousands RSIP Vision has a very long experience in solving these challenges. We have built a unique data augmentation system, able to generate additional CT scans, which look like real even when the anatomy has been deeply modified by Ilya Kovler