Computer Vision News - April 2016

Challenge Ischemic Stroke Lesion Segmentation Every month, Computer Vision News reviews a challenge related to our field, be it in medical imaging, automotive, robotics or else. If you don’t find time to read challenges, but are interested in the new methods proposed by the scientific community to solve them, you can read our challenge summaries! This month we have chosen to review the Ischemic Stroke Lesion Segmentation (ISLES) , a medical image segmentation challenge comparing procedures used on multi-spectral MRI images. The challenge’s website is here . COMPUTER VISION NEWS 20 Background Every year, 15 million people worldwide suffer a stroke . About 6 million die and 5 million suffer a permanent disability. Stroke is the second leading cause of disability, after dementia. Disability may include loss of vision, loss of speech, paralysis, confusion. It drives high social and economic costs. Globally, stroke is the second leading cause of death above the age of 60 years and the fifth leading cause of death in people aged 15 to 59 years. Its most common appearance is ischemic stroke, driven by local thrombosis which is better located, segmented and quantified by analysing brain Magnetic Resonance (MR) scans. Automatic stroke segmentation helps medical experts identify tissue that can be potentially rescued and separate tissue at risk. Time is crucial: a stroke blocking an artery may disrupt the supply of oxygen to brain regions, leading to neuronal death. When blood flow is restored fast enough, parts of the brain tissue may survive and eventually recover. Automatic detection of the rescuable tissue can accelerate the decision making and patient treatment, limiting damage and improving the outcome. Manual segmentation can be performed by professional neuro-radiologists, but it is a time-consuming task and it is non-reproducible. Temporoparietal brain haemorrhage on an MRI scan Suggested Methods and Results The most successful participants suggested a segmentation procedure based on 3D Convolutional Neural Networks , which was able to provide a far better performance than the adaptation of 2D CNNs for processing 3D volumes. The originality of the suggested 11-layer networks system lays in that it was able to train a deep and wide 3D CNN on a very small dataset to complete a segmentation in as fast as 3 minutes. Other high-performing methods employed techniques such as Random Forest classification improved with contextual clustering; some used a modified Random Forest algorithm , adapted for segmentation of medical images and leading to very reliable results in about 6 minutes per every new case. ISLES 2016 has just been launched and the awards of this new challenge will be presented in October.

RkJQdWJsaXNoZXIy NTc3NzU=