Image Segmentation of Brain MRI

This involves training a Unet Model to segment and predict masks from Brain MRI images.

Data: https://www.kaggle.com/mateuszbuda/lgg-mri-segmentation

UNet model architecture:

1. Input - 128X128X1
2. Downward Layers

• 2x Conv Layer - 128x128x64 (D1)
• Max Pool - 64x64x64
• Dropout
• 2x Conv Layer - 64x64x128 (D2)
• Max Pool - 32x32x128
• Dropout
• 2x Conv Layer - 32x32x256 (D3)
• Max Pool - 16x16x256
• Dropout
• 2x Conv Layer - 16x16x512 (D4)
• Max Pool -8x8x512
• Dropout
• 2x Conv Layer - 8x8x1024

3. Upward Layers

• Conv Transpose - 16x16x1024 (U1)
• Concatenate U1 + D4
• Dropout
• 2x Conv Layer - 16x16x1024
• Conv Transpose - 32x32x512 (U2)
• Concatenate U2 + D3
• Dropout
• 2x Conv Layer - 32x32x512
• Conv Transpose - 64x64x256 (U3)
• Concatenate U3 + D2
• Dropout
• 2x Conv Layer - 64x64x256
• Conv Transpose - 128x128x128 (U4)
• Concatenate U4 + D1
• Dropout
• 2x Conv Layer - 128x128x128

4. Prediction layer using Sigmoid

• Conv Layer - 128x128x1