CNN cell detection

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

Getting started

Look at main.ipynb.

Requirements

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

pip3 install -r requirements.txt

Details

• Each image is manually annotated with the center point of each cell as well as some hard negative examples
• All points within sample radius of a cell centre are sampled as positive samples
• An equal number of negative samples are randomly sampled outside the positive radius
• All points within sample radius of the hard negative examples are sampled as negative samples
• A convolutional neural network is trained using the negative and positive samples. For each sample, a box of size box_size, is used as input to the network.
• Given a new image a box_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 possible box_size/2 pixels from the image boundary.

Credit

The network and code structure is based on Lasanges MNIST example https://github.com/Lasagne/Lasagne/blob/master/examples/mnist.py