Daily CVPR - Tuesday

CVPR Daily: Can you explain what kind of algorithm you used to solve this problem? Mahdieh: This pipeline has actually two main steps. One is segmentation. In the segmentation part, we used a novel and complicated segmentation algorithm which is called C-EPAC. This segmentation part is novel, and is pretty accurate to split the clumped cells. Clumped cells means cells which are having overlap together. So for example, you have single cells vs. there are different cells which are overlapped with each other. So if you want to figure out if the cells are infected or not, first of all, you need to split them and find the individual cells. So this is a really challenging task: how to split clumped cells. And this method is using a four color level set segmentation approach to address this clump exposing issue. CVPR Daily: What difficulties did you find with this kind of algorithm? Mahdieh: With this problem, the main issue is about its timing. Since it goes through different iterations to achieve higher accuracy, that is also really time consuming. So when your image is really big, it takes almost 5 minutes on a machine to extract or to create, generate segmentation mask for you. And this really a kind of thing we are facing since we are going to develop a mobile app for this project. And on the app, you cannot go for such a time consuming algorithm. So it’s pretty good. The results are really nice. You can see we got 98% precision for one of our datasets for mouse and 97% for human images. This is really accurate, but if you are going to use it in the field on the mobile app, it needs to be really quick and fast. CVPR Daily: What is the next step for this research? Mahdieh: For this research, the next step is to find a kind of reasonable, fast, and accurate segmentation algorithm. And also at some point, you are going to apply deep learning which is state of the art these days, to first compare our results to deep learning methods, and also see how we can take the advantage of deep learning for this project. CVPR Daily: Tuesday Presentation 9