Скачать или смотреть Comparing Image Segmentation Methods for Nucleus Detection

About this project: Segmenting the nuclei of cells in microscopy images is critical for many biomedical applications; by measuring how cells react to various treatments, researchers can understand the underlying biological processes at work. Therefore, automating nucleus detection could help unlock cures faster. In this study, we reviewed several deep learning architectures for image segmentation. We compared the performance of various U-Net based architectures, DeepLab network, and modified U-Net with various backbone; to perform segmentation of nucleus images from the 2018 Data Science Bowl competition dataset. Our results show that these U-Net based networks outperformed ImageNet pretrained networks. These models were able to generalize to different lighting conditions and nucleus type. As a result, minimal tuning is needed for different nucleus type; this is desirable in clinical settings. Source code can be found at https://github.com/jinglescode/meditorch.

Assignment coursework for DM6190 in School of Computer Science and Engineering, Nanyang Technological University