Final test rand score: 0.979
The architecture was inspired by U-Net: Convolutional Networks for Biomedical Image Segmentation and implemented with Keras functional API.
ISBI 2012 dataset: http://brainiac2.mit.edu/isbi_challenge/
The dataset is in folder data
.
training hyper-parameters:
- checkpoint_name: save_dir of tensorboard, set experiment name
- batch_size: batch size, integer multiple of GPU numbers
- prop_trn: proportion of training data 20./30
- prop_val: proportion of validation data
- montage_trn_shape: width and height of training data combination (5, 4) 5*4=20
- montage_val_shape: width and height of validation data combination (5, 2) 5*2=10
- early_stop_patience: 30
- aug: whether to use deformation in data augmentation
- steps: steps of each epoch
- random_split: whether to random split the 30 images data
Loss functions src/utils/model.py
:
- weighted cross entropy
- focal loss
- dice loss
- tversky loss
- combinations
train on the isbi training data
predict on validation or test data
check the results after different data augmentation