232 - Semantic Segmentation of BraTS2020 - Part 1 - Getting the data ready

Code generated in the video can be downloaded from here:
https://github.com/bnsreenu/python_fo...

Dataset from: https://www.kaggle.com/awsaf49/brats2...

Dataset information:
Multimodal scans available as NIfTI files (.nii.gz)

Four 'channels' of information – 4 different volumes of the same region
Native (T1)
Post-contrast T1-weighted (T1CE)
T2-weighted (T2)
T2 Fluid Attenuated Inversion Recovery (FLAIR) volumes

All the imaging datasets have been segmented manually and were approved by experienced neuro-radiologists.

Annotations (labels):
Label 0: Unlabeled volume
Label 1: Necrotic and non-enhancing tumor core (NCR/NET)
Label 2: Peritumoral edema (ED)
Label 3: Missing (No pixels in all the volumes contain label 3)
Label 4: GD-enhancing tumor (ET)

Our approach:
Step 1: Get the data ready (this video)
Step 2: Define custom data generator (next video)
Step 3: Define the 3D U-net model
Step 4: Train and Predict

Step 1: Get the data ready
Download the dataset and unzip it.
Segmented file name in Folder 355 has a weird name. Rename it to match others.
Install nibabel library to handle nii files (https://pypi.org/project/nibabel/)
Scale all volumes (using MinMaxScaler).
Combine the three non-native volumes (T2, T1CE and Flair) into a single multi-channel volume.
Reassign pixels of value 4 to value 3 (as 3 is missing from original labels).
Crop volumes to remove useless blank regions around the actual volume of interest (Crop to 128x128x128).
Drop all volumes where the amount of annotated data is less that certain percentage. (To maximize training on real labeled volumes).
Save all useful volumes to the local drive as numpy arrays (npy).
Split image and mask volumes into train and validation datasets.