Brain Segmentation on MRBrains18
Classes | label |
---|---|
Cortical gray matter | 1 |
Basal ganglia | 2 |
White matter | 3 |
White matter lesions | 4 |
Cerebrospinal fluid in the extracerebral space | 5 |
Ventricles | 6 |
Cerebellum | 7 |
Brain stem | 8 |
Regularized Biased Field Corrected MRI | Removed Skull | Histogram equalization |
Cortical gray matter, White matter, Cerebrospinal fluid in the extracerebral space can be easily reduced by appling thresholding to T1- weighted MRI further a small U-Net was used to denoise the threshold image.
Total parameters: 60,553  For rest of the classes training was done on a custom model inspired by Unet Architecture. The model has 3 encoders stacked together in bottleneck layer and then a single decoder. There are skip connections from encoder to decoder to enhance segmentation.
Total parameters: 151,717U-Net | Architecture used |
---|---|
Only one encoder and one decoder | Three encoder and one decoder |
Deep architecture with about 10 Million parameters | Shallow with about 600 Thousands parameters |
Doesn’t have dilated convolution layers | Has dilated convolution layers |
Dice coefficient is used as Loss function in final training though Jaccard distance and crossentropy were also tried.
Label | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
---|---|---|---|---|---|---|---|---|
Dice coefficient | 0.702 | 0.758 | 0.770 | 0.746 | 0.704 | 0.882 | 0.887 | 0.855 |
This project was made as part of the Smart India hackathon 2018 - Software Edition, a 36 hour hackathon organised by Government of India. The problem statement was given by Department of Atomic Energy, India