Skip to content

CNN Predictions

alt text

The CNN predictions widget process the stacks at hand with a Convolutional Neural Network. The output is a boundary classification image, where every voxel gets a value between 0 (not a cell boundary) and 1 (cell boundary).

The input image can be a raw stack "tiff"/"h5" or the output of the PreProcessing widget.

  • The Model Name menu shows all available models. There are two main basic models available

    1. Generic confocalis a generic model for all confocal datasets. Some examples: alt text

    2. Generic lightsheet this is a generic model for all lightsheet datasets. Some examples: alt text

  • Due to memory constraints, usually a complete stack does not fit the GPUs memory, therefore the Patch size can be used to optimize the performance of the pipeline. Usually, larger patches cost more memory but can slightly improve performance. For 2D segmentation, the Patch size relative to the z-axis has to be set to 1.

  • To minimize the boundary effect due to the sliding windows patching, we can use different stride:

    1. Accurate: corresponding to a stride 50% of the patch size (yield best predictions/segmentation accuracy)
    2. Balanced: corresponding to a stride 75% of the patch size
    3. Draft: corresponding to a stride 95% of the patch size (yield fastest runtime)

  • The Device type menu can be used to enable or not GPU acceleration. CUDA greatly accelerates the network predictions on Nvidia GPUs. At the moment, we don't support other GPUs manufacturers.