yapic_io

yapic_io provides flexible data binding to image collections of arbitrary size.

Its aim is to provide a convenient image data interface for training of fully convolutional neural networks, as well as automatic handling of prediction data output for a trained classifier.

yapic_io is designed as a convenient image data input/output interface for libraries such as Theano or TensorFlow.

Following problems occuring with training/classification are handeled by yapic_io:

• Images of different sizes in z,x, and y can be applied to the same convolutional network. This is implemented by sliding windows. The size these windows correspond to the size of the convolutional network's input layer.

• Due to lazy data loading, images can be extremely large.

• Image dimensions can be up to 4D (multi-channel z-stack), as e.g. required for bioimages.

• Data augmentation for classifier training in built in.

• Made for sparsly labelled datasets: Training data is only (randomly) picked from regions where labels are present.

• Usually, input layers of CNNs are larger than output layers. Thus, pixels located at image edges are normally not classified. With yapic_io also edge pixels are classified. This is achieved by mirroring pixel data in edge regions. As a result, output classification images have identical dimensions as source images and can be overlayed easily.

Currently supported image annotation platforms

• napari
• ilastik

Example Classifier

Training:

from yapic_io import TiffConnector, Dataset, TrainingBatch

#define data locations
pixel_image_dir = 'yapic_io/test_data/tiffconnector_1/im/*.tif'
label_image_dir = 'yapic_io/test_data/tiffconnector_1/labels/*.tif'
savepath = 'yapic_io/test_data/tmp/'


tpl_size = (1,5,4) # size of network output layer in zxy
padding = (0,2,2) # padding of network input layer in zxy, in respect to output layer

c = TiffConnector(pixel_image_dir, label_image_dir, savepath=savepath)
train_data = TrainingBatch(Dataset(c), tpl_size, padding_zxy=padding)

counter=0
for mini in train_data:
     weights = mini.weights
     #shape of weights is (6,3,1,5,4) : batchsize 6 , 3 label-classes, 1 z, 5 x, 4 y

     pixels = mini.pixels()
     # shape of pixels is (6,3,1,9,8) : 3 channels, 1 z, 9 x, 4 y (more xy due to padding)

     #here: apply training on mini.pixels and mini.weights (use theano, tensorflow...)
     my_train_function(pixels, weights)

     counter += 1
     if counter > 10: #m is infinite
         break

Prediction:

from yapic_io import TiffConnector, Dataset, PredictionBatch

#mock classification function
def classify(pixels, value):
    return np.ones(pixels.shape) * value

#define data loacations
pixel_image_dir = 'yapic_io/test_data/tiffconnector_1/im/*.tif'
label_image_dir = 'yapic_io/test_data/tiffconnector_1/labels/*.tif'
savepath = 'yapic_io/test_data/tmp/'

tpl_size = (1,5,4) # size of network output layer in zxy
padding = (0,2,2) # padding of network input layer in zxy, in respect to output layer

c = TiffConnector(pixel_image_dir, label_image_dir, savepath=savepath)
prediction_data = PredictionBatch(Dataset(c))
print(len(prediction_data)) #give the total number of templates that cover the whole bound tifffiles

#classify the whole bound dataset
counter = 0 #needed for mock data
for item in prediction_data:
    pixels_for_classifier = item.pixels() #input for classifier
    mock_classifier_result = classify(pixels, counter) #classifier output

    #pass classifier results for each class to data source
    item.put_probmap_data(mock_classifier_result)

    counter += 1 #counter for generation of mockdata

Buils API docs

cd docs
sphinx-apidoc -o source ../yapic_io
make html

Developed by the CRFS (Core Research Facilities) of the DZNE (German Center for Neurodegenerative Diseases).