MohamadZeina / 3D-Patch-Based-Keras-Unet

A modular, 3D unet built in keras for 3D medical image segmentation. Also includes useful classes for extracting and training on 3D patches for data augmentation or memory efficiency.

3D_patch_processing_utils

About

Useful utilities for working with 3D data in patches. These were designed to train and evaluate deep learning models for 3D segmentation of brain MRI data. Main classes:

• CategoriseNiftis: this takes a list of niftis, and corresponding SPM segmentations and makes it easy to access raw data and their corresponding segmentation files
• PatchSequence: this is a keras generator, which inherits the "Sequence" class. This takes a list of file paths, and returns shuffled patches for training.
• UnetEvaluator: this overrides PatchSequence. It contains various methods for evaluating models. For example, it might take a model and an unsegmented volume, segment it and display the output.

Example Usage

niftis_path = "/path/to/images" # Points to images which have been segmented in SPM 

model = your_keras_model()

niftis = CategoriseNiftis(niftis_path, require_oasis=False, require_string='T1')

generator = PatchSequence(
    [niftis.raw], [niftis.seg_1, niftis.seg_2, niftis.seg_3], 
    batch_size=16, patch_size=128, stride = 64)
    
history = model.fit_generator(generator, max_queue_size=200, shuffle = False)

Acknowledgements

Data used during development, and in above visualisation, by OASIS:

• OASIS-3: Principal Investigators: T. Benzinger, D. Marcus, J. Morris; NIH P50AG00561, P30NS09857781, P01AG026276, P01AG003991, R01AG043434, UL1TR000448, R01EB009352. AV-45 doses were provided by Avid Radiopharmaceuticals, a wholly owned subsidiary of Eli Lilly.