Repository files navigation

This code show how to train a cell detector using a convolutional neural network inLasagne.

Look atmain.ipynb.

• Python 2 or 3
• The python packages inrequirements.txt, if you have pip you can install them using:

pip3 install -r requirements.txt

• Each image is manually annotated with the center point of each cell as well as some hard negative examples
• All points withinsample radius of a cell centre are sampled aspositive samples
• An equal number ofnegative samples are randomly sampled outside thepositive radius
• All points withinsample radius of the hard negative examples are sampled asnegative samples
• Aconvolutional neural network is trained using the negative and positive samples. For each sample, a box of sizebox_size, is used as input to the network.
• Given a new image abox_sized window is slided through each possible patch in the image, generating a probability map
• Local maxima in the probability map are marked as cell centers

Note: There is no padding on the boundary so no detection is possiblebox_size/2 pixels from the image boundary.

The network and code structure is based on LasangesMNIST examplehttps://github.com/Lasagne/Lasagne/blob/master/examples/mnist.py