Computer Vision News - March 2023

45 Synapses Detection d so that some patches might not contain any T-bar. ntral point is used as a channel of input patch. Note that isotropic size. “ I’m not a specialist in this field, so I hope they will try their best to solve our problem, ” he appeals. “ This is a real problem I encountered in my research. I’m working on a microwasp connectomics project to map several microwasp brains . It’s a rare study area, and to the best of my knowledge, we’re the only organization doing it. We have three data sets covering the whole brain, and I work mostly on one of them. I’ve built a baseline method, but it doesn’t generalize well across brain regions and samples. ” The wider connectomics community is experiencing rapid growth, and Jingpeng’s work extends beyond microwasps to include flies, mice, monkeys, and even humans . However, the methods used in mammalian brains cannot be directly applied to microwasps. One of the most significant differences is in the post- synapse density or synaptic cleft . “ In a mammalian brain, you can see a very thick membrane when two neurons meet with a synapse , ” he explains. “ The contacting face of the synapse is very dense, dark, and thick and is called Post Synapse Density. In microwasps, the synapses are much smaller, making it harder to see this post-synapse density. Also, our imaging method has a lower lateral resolution than transmission electron microscopy, making this problem more difficult to solve than in a mammalian brain. ” Youcan learnmoreabout this competition on the challenge website, and Jingpeng’s full proposal is available on arXiv. cause images of the same species, sex, and age to appear noticeably different. “ We want to build an atlas of the brain , ” Jingpeng reveals. “ This requires very high accuracy. We’ll approach humans as proofreaders to correct errors, but there’s a balance between howmuch human work is needed and how accurate the model must be. Improving these models will gradually reduce the work required by humans. ” He hopes the challenge attracts computer vision community members interested in out-of-domain object detection, often experts at building models with good generalizability. ISBI 2023 CHALLENGE Ground truth Ground truth Postsynapse prediction Presynapse prediction es re randomly s mpled so hat some patches might er of each patch and a fixed patch with a central p int d network are 3D and of isotropic size.