deep-review
deep-review copied to clipboard
greenelab
U-Net: deep learning for cell counting, detection, and morphometry
U-Net is a generic deep-learning solution for frequently occurring quantification tasks such as cell detection and shape measurements in biomedical image data. We present an ImageJ plugin that enables non-machine-learning experts to analyze their data with U-Net on either a local computer or a remote server/cloud service. The plugin comes with pretrained models for single-cell segmentation and allows for U-Net to be adapted to new tasks on the basis of a few annotated samples.
https://doi.org/10.1038/s41592-018-0261-2
It doesn't look like we talk about U-Net in this repo yet, but I came across this recent brief communication today.
Summary
The authors present an ImageJ plugin of the U-Net architecture with pretrained weights. The package is used as a "generic" deep learning solution that can adapt to various cell detection and segmentation tasks across imaging domains.
One limitation is the requirement for users to fine-tune the weights of the U-Net model with their own labeled data. While this adds the benefit of customized solutions within labs, and, presumably, the software will get better with use, how far can the software diverge across labs? If two labs use the plugin for a year and then analyze the same image will they get different results?
Nevertheless, I thought it was a cool application of democratizing deep learning to non computational scientists. A user experienced with ImageJ can use this software with pretrained weights. The authors note that this feature is similar to other packages Aivia and CellProfiler.
@gwaygenomics Thanks for the issue and the summary. There's a variant of U-Net, the Probabilistic U-Net (Kohl et al. 2018), that combines a CVAE with a U-Net for conditional density estimation over segmentations. They use the LIDC-IDRI dataset of manual lesion segmentations from lung patients to assess their framework but it seems ancillary to the discussion of the architecture (they also test on the Cityscapes dataset as a segmentation task). Do you think this might be worth including as a separate issue?
