Code and data used in publication Accurate and robust segmentation of neuroanatomy in T1-weighted MRI by combining spatial priors with deep convolutional neural networks

Prerequisites

• Tested on Python 3.6.3
• MINC toolkit
• Pyminc
• Lasagne + Theano
• Scipy

Example code

Train a network:

python train_cnn.py -tr_ima yourTrainingImages.txt -tr_mask yourTrainingMasks.txt -model_name yourModelName -sampling_mask yourSamplingMask.mnc -num_labels yourLabelNumber

Apply it:

python apply_cnn.py -te_ima yourTrainingImages.txt -output_dir yourOutputDir -model_name yourModelName -sampling_mask yourSamplingMask.mnc -num_labels yourLabelNumber -positive_mask yourPositiveMask.mnc

Details

Images should be intensity-normalized and registered in a common space. Make sure your label images are integer-valued (background should have a value of 0). See paper and code comments for further details.