Fully Convolutional Network Architecture for the Automation of MR Image Segmentation

load data

• load_dicom
  • loads all the .dcm files in a filepath into a np.ndarray
• load_nrrd
  • loads a single .nrrd file from filepath into a np.ndarray
  • optionally returns the header information

Network Architecture and Ideas:

• Fully Convolutional Network
• Mark regions that are somewhat uncertain
  • Use Seg3D2 or own simple verification application
  • Different mask for each tissue + different mask for uncertain tissues
  • Re-trainable with new data -> after manual verification

Issues:

• Output format?
• Prevent Catastrophic Forgetting
• Keep lr python currentLearningRate = K.get_value(model.optimizer.lr)