Semantic Segmentation for Medical Images

Implemenation of Unets for Lung Segmentation in Xrays in Keras

Key Details

• Implementation of a Unet with Keras
• Based on the work U-Net: Convolutional Networks for Biomedical Image Segmentation
• Segmented dataset for training obtained from TB Chest Xray Image Sets

Item	Details
Input	256 x 256 grayscale Xray Image
Output	256 x 256 segmentation map
Train Images	110
Manual train masks	110
Validation Images	28
Manual validation masks	28

• Thanks to zhixuhao for the keras implementation of unets
• Have improved upon that to run with image generators in keras dynamically and augment while training

---

Dependencies

• Keras 2.1.5
• Numpy 1.14.2
• OpenCV 2.4.9.1
  • Just using it to write and resize images
  • You may replace with PIL if you prefer

---

Things to note

• While running ensure that the xrays and images are in separate folders and have the same labels
• Follow similar folder hierarchy in data/ to your work easier ;)

---

Running Unets

# Initialize the Unet
u1 = Unet()

# Round one of training
u1.train(lr=1e-4,num_epochs=20)

# Improve upon existing model
u1.continue_training(lr=1e-4,num_epochs=20)

# Visualize image and output side by side
u1.generate_output(save=True,mode='side_by_side',output_folder='data/outputs/side_by_side/')

# Crop images based on output mask and return the mask
u1.generate_output(save=True,mode='cropped',output_folder='data/outputs/cropped/')

# Get just the masks
u1.generate_output(save=True,mode='mask_only',output_folder='data/outputs/masks_only/')