The data is supplied by CBICA at Perelman School of Medicine, University of Pennsylvania and is avaiable on Kaggle. The data are multimodal scans and are given as .nii.gz files. The datset contains 4 volumes but in this project one three volumes are used: the post-contrast T1-weighted (T1Gd), T2-weighted (T2) and Fluid Attenuated Inversion Recovery (T2-FLAIR).
Slice of a brain volume:
Preprocessing steps include Min-Max-Scaling, converting nii files to numpy arrays and cropping.
For the purposes of this project, the volume is cropped into the following format:
After scaling:
Training a simple Unet model adapted to the 3D setting, consisting of 5,645,828 parameters and is trained for 100 epochs.
A 3D Unet with attention blocks is also trained for 100 epochs.
Comparing the mean IoU on the test data for the two models. It seems that the attention model is slightly better compared to the base model.
Basic 3D Unet: 0.7957 Attention 3D Unet: 0.8339
