- I’m having trouble with DeepCell. Can I get help.
- Yes. If you think you have discovered a bug that needs to be fixed please file a report on the SimTK page.
- What kind of hardware do I need to run DeepCell?
- You will need a CUDA/cuDNN capable Nvidia GPU. We have had good success with the Nvidia GTX 980, Titan X, and GTX 1080 graphics cards.
- Does DeepCell work with TensorFlow?
- Unfortunately no. We have found that TensorFlow is unable to use numpy like indexing to address tensors. This makes it significantly harder to implement d-regularly sparse pooling kernels - but we’re working on it.
- Does DeepCell track cells from frame to frame?
- Unfortunately no. Right now, this software package focuses solely on the image segmentation problem for live-cell experiments. However, we are aware that cell tracking is an issue for a number of labs and we’re actively working on deep learning approaches to this problem.
- What cells can DeepCell segment?
- So far we have trained convolutional neural networks to segment fluorescently labeled nuclei, as well as phase images of E. coli, MCF10A cells, NIH-3T3 cells, HeLa-S3 cells, RAW264.7 cells, and bone marrow derived macrophages.
- Do I need a nuclear label to segment the cytoplasm of mammalian cells?
- For our approach, yes. The nuclear labels are necessary to refine the segmentation prediction.
- Should I train my own network?
- We recommend it. There are laboratory-to-laboratory differences (lighting, microscope, camera, pixel size, etc.) that do matter.
- Where should I go if I want to learn more about deep learning?