U-Net with a CRF-RNN layer

This project aims at improving U-Net for medical images segmentation. Our model was implemented using Tensorflow and Keras, and the CRF-RNN layer refers to this repo

Introducion

• U-Net with CRF-RNN layer paper:

1. UNet-CRF-RNN

• Reference paper:

1. U-Net
2. FCN
3. CRF-RNN

This repo provides an U-Net with the CRF-RNN layer, and also provides some extract models for comparison, like SegNet, FCN, vanilla U-Net and so on.

modelFns = {'unet':Models.VanillaUnet.VanillaUnet, 
            'segnet':Models.Segnet.Segnet , 
            'vgg_unet':Models.VGGUnet.VGGUnet , 
            'vgg_unet2':Models.VGGUnet.VGGUnet2 , 
            'fcn8':Models.FCN8.FCN8, 
            'fcn32':Models.FCN32.FCN32, 
            'crfunet':Models.CRFunet.CRFunet   }

Usage

• data hierarchy

    Use the Keras data generators to load train and test
    Image and label are in structure:
        train/
            img/
                0/
            gt/
                0/

        test/
            img/
                0/
            gt/
                0/

• Training parameters

'--batch_size', type=int, default=1, help='input batch size'
'--learning_rate', type=float, default=0.0001, help='learning rate'
'--lr_decay', type=float, default=0.9, help='learning rate decay'
'--epoch', type=int, default=80, help='# of epochs'
'--imSize', type=int, default=320, help='then crop to this size'
'--iter_epoch', type=int, default=0, help='# of iteration as an epoch'
'--num_class', type=int, default=2, help='# of classes'
'--checkpoint_path', type=str, default='', help='where checkpoint saved'
'--data_path', type=str, default='', help='where dataset saved. See loader.py to know how to organize the dataset folder'
'--load_from_checkpoint', type=str, default='', help='where checkpoint saved'

• Train your model

  python train.py --data_path ./datasets/ --checkpoint_path ./checkpoints/
  

• Visualize the train loss, dice score, learning rate, output mask, and first layer convolutional kernels per iteration in tensorboard

  tensorboard tensorboard --logdir=./checkpoints
  

• Evaluate your model

  python eval.py --data_path ./datasets/ --load_from_checkpoint ./checkpoints/model-xxxx
  

Result

• Dataset

1. Hippocampus Segmentation: ADNI
2. Hippocampus Segmentation: NITRC

• Parameters

• Result