231 - Semantic Segmentation of BraTS2020 - Part 0 - Introduction (and plan)

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
Step 2: Define custom data generator
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. 

Step 2: Define custom data generator
Keras image data generator only works with jpg, png, and tif images. It will not recognize npy files. We need to define a custom generator to load our data from the disk. 

Step 3: Define the 3D U-net model
Extend the standard 2D U-Net into 3D OR
copy the code from online OR
use 3D segmentation models library 

Step 4: Train and Predict
Train by loading images in batches using our custom generator. 
Predict and plot data for visualization.