Merge pull request #11 from HiLab-git/dev

Dev

Author: taigw

.gitignore

@@ -1,5 +1,5 @@
-PyMIC_data/ACDC/*
-PyMIC_data/Fetal_HC/*
-PyMIC_data/JSRT/*
-PyMIC_data/MyoPS/*
-PyMIC_data/Promis12/*
+PyMIC_data/ACDC/
+PyMIC_data/Fetal_HC
+PyMIC_data/JSRT
+PyMIC_data/MyoPS
+PyMIC_data/Promise12

README.md

@@ -5,7 +5,7 @@
 The latest released version of PyMIC can be installed by:
 
 ```bash
-pip install PYMIC
+pip install PYMIC==0.3.1.1
 ```
 
 To use the latest development version, you can download the source code [here][PyMIC_link], and install it by:
@@ -29,7 +29,8 @@ Currently we provide the following examples in this repository:
 |Fully supervised segmentation|[JSRT2][JSRT2_link]|Using a customized network and loss function for the JSRT dataset|
 |Fully supervised segmentation|[Fetal_HC][fetal_hc_link]|Using a 2D UNet for fetal head segmentation from 2D ultrasound images|
 |Fully supervised segmentation|[Prostate][prostate_link]|Using a 3D UNet for prostate segmentation from 3D MRI|
-|Semi-supervised segmentation|[seg_ssl/ACDC][ssl_acdc_link]|Comparing different semi-supervised methods for heart structure segmentation|
+|Semi-supervised segmentation|[seg_ssl/ACDC][ssl_acdc_link]|Semi-supervised methods for heart structure segmentation using 2D CNNs|
+|Semi-supervised segmentation|[seg_ssl/AtriaSeg][ssl_atrial_link]|Semi-supervised methods for left atrial segmentation using 3D CNNs|
 |Weakly-supervised segmentation|[seg_wsl/ACDC][wsl_acdc_link]|Segmentation of heart structure with scrible annotations|
 |Noisy label learning|[seg_nll/JSRT][nll_jsrt_link]|Comparing different NLL methods for learning from noisy labels|
 
@@ -43,6 +44,7 @@ Currently we provide the following examples in this repository:
 [fetal_hc_link]:segmentation/fetal_hc
 [prostate_link]:segmentation/prostate
 [ssl_acdc_link]:seg_ssl/ACDC
+[ssl_atrial_link]:seg_ssl/AtriaSeg/
 [wsl_acdc_link]:seg_wsl/ACDC 
 [nll_jsrt_link]:seg_nll/JSRT
 

classification/AntBee/README.md

@@ -13,14 +13,14 @@ In this example, we finetune a pretrained resnet18 for classification of images
 [data_link]:https://download.pytorch.org/tutorial/hymenoptera_data.zip
 
 ## Finetuning all layers of resnet18
-1. Here we use resnet18 for finetuning, and update all the layers. Open the configure file `config/train_test_ce1.cfg`. In  the `network` section we can find details for the network. Here `update_layers = 0` means updating all the layers.
+1. Here we use resnet18 for finetuning, and update all the layers. Open the configure file `config/train_test_ce1.cfg`. In  the `network` section we can find details for the network. Here `update_mode = all` means updating all the layers.
 ```bash
 # type of network
 net_type   = resnet18
 pretrain   = True
 input_chns = 3
 # finetune all the layers
-update_layers = 0  
+update_mode = all 
 ```
 
 Then start to train by running:
@@ -48,20 +48,20 @@ pymic_run test config/train_test_ce1.cfg
 pymic_eval_cls config/evaluation.cfg
 ```
 
-The obtained accuracy by default setting should be around 0.9412, and the AUC will be around 0.976.
+The obtained accuracy by default setting should be around 0.9477, and the AUC will be around 0.9745.
 
 3. Run `python show_roc.py` to show the receiver operating characteristic curve. 
 
 ![roc](./picture/roc.png)
 
 ## Finetuning the last layer of resnet18
-Similarly to the above example, we further try to only finetune the last layer of resnet18 for the same classification task. Use a different configure file `config/train_test_ce2.cfg` for training and testing, where `update_layers = -1` in the `network` section means updating the last layer only:
+Similarly to the above example, we further try to only finetune the last layer of resnet18 for the same classification task. Use a different configure file `config/train_test_ce2.cfg` for training and testing, where `update_mode = last` in the `network` section means updating the last layer only:
 ```bash
 net_type   = resnet18
 pretrain   = True
 input_chns = 3
 # finetune the last layer only
-update_layers = -1
+update_mode = last
 ```
 
-Edit `config/evaluation.cfg` accordinly for evaluation. 
+Edit `config/evaluation.cfg` accordinly for evaluation. The corresponding accuracy and AUC would be around 0.9477 and 0.9778, respectively.

classification/AntBee/config/train_test_ce1.cfg

@@ -14,8 +14,8 @@ train_batch_size = 4
 modal_num = 1
 
 # data transforms
-train_transform = [Rescale, RandomCrop, RandomFlip, NormalizeWithMeanStd]
-valid_transform = [Rescale, CenterCrop, NormalizeWithMeanStd]
+train_transform = [Rescale, RandomCrop, RandomFlip, NormalizeWithMeanStd, LabelToProbability]
+valid_transform = [Rescale, CenterCrop, NormalizeWithMeanStd, LabelToProbability]
 test_transform  = [Rescale, CenterCrop, NormalizeWithMeanStd]
 
 Rescale_output_size    = [256, 256]
@@ -39,7 +39,7 @@ net_type   = resnet18
 pretrain   = True
 input_chns = 3
 # finetune all the layers
-update_layers = 0  
+update_mode = all
 
 # number of classes
 class_num = 2
@@ -56,19 +56,20 @@ learning_rate = 1e-3
 momentum      = 0.9
 weight_decay  = 1e-5
 
-# for lr schedular (MultiStepLR)
-lr_scheduler  = MultiStepLR
-lr_gamma      = 0.1
-lr_milestones = [500, 1000]
+# for lr schedular (StepLR)
+lr_scheduler = StepLR
+lr_gamma     = 0.5
+lr_step      = 500
 
-ckpt_save_dir    = model/resnet18_ce1
-ckpt_prefix = resnet18
+ckpt_save_dir = model/resnet18_ce1
+ckpt_prefix   = resnet18
 
 # iteration
 iter_start = 0
-iter_max   = 1500
+iter_max   = 2000
 iter_valid = 100
-iter_save  = 1500
+iter_save  = 2000
+early_stop_patience = 1000
 
 [testing]
 # list of gpus

classification/AntBee/config/train_test_ce2.cfg

@@ -14,8 +14,8 @@ train_batch_size = 4
 modal_num = 1
 
 # data transforms
-train_transform = [Rescale, RandomCrop, RandomFlip, NormalizeWithMeanStd]
-valid_transform = [Rescale, CenterCrop, NormalizeWithMeanStd]
+train_transform = [Rescale, RandomCrop, RandomFlip, NormalizeWithMeanStd, LabelToProbability]
+valid_transform = [Rescale, CenterCrop, NormalizeWithMeanStd, LabelToProbability]
 test_transform  = [Rescale, CenterCrop, NormalizeWithMeanStd]
 
 Rescale_output_size    = [256, 256]
@@ -39,8 +39,7 @@ net_type   = resnet18
 pretrain   = True
 input_chns = 3
 # finetune the last layer only
-update_layers = -1
-
+update_mode = last
 
 # number of classes
 class_num = 2
@@ -57,19 +56,20 @@ learning_rate = 1e-3
 momentum      = 0.9
 weight_decay  = 1e-5
 
-# for lr schedular (MultiStepLR)
-lr_scheduler  = MultiStepLR
-lr_gamma      = 0.1
-lr_milestones = [500, 1000]
+# for lr schedular (StepLR)
+lr_scheduler = StepLR
+lr_gamma     = 0.5
+lr_step      = 500
 
 ckpt_save_dir    = model/resnet18_ce2
 ckpt_prefix = resnet18
 
 # iteration
 iter_start = 0
-iter_max   = 1500
+iter_max   = 2000
 iter_valid = 100
-iter_save  = 1500
+iter_save  = 2000
+early_stop_patience = 1000
 
 [testing]
 # list of gpus

classification/CHNCXR/README.md

@@ -20,7 +20,7 @@ net_type = resnet18
 pretrain = True
 input_chns = 3
 # finetune all the layers
-update_layers = 0
+update_mode = all
 ```
 
 Start to train by running:
@@ -48,12 +48,12 @@ pymic_run test config/net_resnet18.cfg
 pymic_eval_cls config/evaluation.cfg
 ```
 
-The obtained accuracy by default setting should be around 0.8571, and the AUC is 0.94.
+The obtained accuracy by default setting should be around 0.8271, and the AUC is 0.9343.
 
 3. Run `python show_roc.py` to show the receiver operating characteristic curve. 
 
 ![roc](./picture/roc.png)
 
 
 ## Finetuning vgg16
-Similarly to the above example, we further try to finetune vgg16 for the same classification task. Use a different configure file `config/net_vg16.cfg` for training and testing. Edit `config/evaluation.cfg` accordinly for evaluation. The iteration number for the highest accuracy on the validation set was 2300, and the accuracy will be around 0.8797. 
+Similarly to the above example, we further try to finetune vgg16 for the same classification task. Use a different configure file `config/net_vg16.cfg` for training and testing. Edit `config/evaluation.cfg` accordinly for evaluation. The accuracy and AUC would be around 0.8571 and 0.9271, respectively. 

classification/CHNCXR/config/net_resnet18.cfg

@@ -14,8 +14,8 @@ train_batch_size = 4
 modal_num = 1
 
 # data transforms
-train_transform = [Rescale, RandomCrop, RandomFlip, GrayscaleToRGB, NormalizeWithMeanStd]
-valid_transform = [Rescale, CenterCrop, GrayscaleToRGB, NormalizeWithMeanStd]
+train_transform = [Rescale, RandomCrop, RandomFlip, GrayscaleToRGB, NormalizeWithMeanStd, LabelToProbability]
+valid_transform = [Rescale, CenterCrop, GrayscaleToRGB, NormalizeWithMeanStd, LabelToProbability]
 test_transform  = [Rescale, CenterCrop, GrayscaleToRGB, NormalizeWithMeanStd]
 
 Rescale_output_size    = [256, 256]
@@ -37,7 +37,7 @@ net_type = resnet18
 pretrain = True
 input_chns = 3
 # finetune all the layers
-update_layers = 0
+update_mode = all
 
 # number of classes
 class_num = 2
@@ -54,10 +54,10 @@ learning_rate = 1e-3
 momentum      = 0.9
 weight_decay  = 1e-5
 
-# for lr schedular (MultiStepLR)
-lr_scheduler  = MultiStepLR
-lr_gamma      = 0.1
-lr_milestones = [1500, 3000]
+# for lr schedular (StepLR)
+lr_scheduler  = StepLR
+lr_gamma      = 0.5
+lr_step       = 1000
 
 ckpt_save_dir    = model/resnet18
 ckpt_prefix      = resnet18
@@ -66,7 +66,8 @@ ckpt_prefix      = resnet18
 iter_start = 0
 iter_max   = 5000
 iter_valid = 100
-iter_save  = 1000
+iter_save  = 5000
+early_stop_patience = 2000
 
 [testing]
 # list of gpus

classification/CHNCXR/config/net_vgg16.cfg

@@ -14,8 +14,8 @@ train_batch_size = 4
 modal_num = 1
 
 # data transforms
-train_transform = [Rescale, RandomCrop, RandomFlip, GrayscaleToRGB, NormalizeWithMeanStd]
-valid_transform = [Rescale, CenterCrop, GrayscaleToRGB, NormalizeWithMeanStd]
+train_transform = [Rescale, RandomCrop, RandomFlip, GrayscaleToRGB, NormalizeWithMeanStd, LabelToProbability]
+valid_transform = [Rescale, CenterCrop, GrayscaleToRGB, NormalizeWithMeanStd, LabelToProbability]
 test_transform  = [Rescale, CenterCrop, GrayscaleToRGB, NormalizeWithMeanStd]
 
 Rescale_output_size    = [256, 256]
@@ -37,7 +37,7 @@ net_type   = vgg16
 pretrain   = True
 input_chns = 3
 # finetune all the layers
-update_layers = 0
+update_mode = all
 
 # number of classes
 class_num = 2
@@ -54,10 +54,10 @@ learning_rate = 1e-3
 momentum      = 0.9
 weight_decay  = 1e-5
 
-# for lr schedular (MultiStepLR)
-lr_scheduler  = MultiStepLR
-lr_gamma      = 0.1
-lr_milestones = [1500, 3000]
+# for lr schedular (StepLR)
+lr_scheduler  = StepLR
+lr_gamma      = 0.5
+lr_step       = 1000
 
 ckpt_save_dir = model/vgg16
 ckpt_prefix   = vgg16
@@ -66,7 +66,8 @@ ckpt_prefix   = vgg16
 iter_start = 0
 iter_max   = 5000
 iter_valid = 100
-iter_save  = 1000
+iter_save  = 5000
+early_stop_patience = 2000
 
 [testing]
 # list of gpus

seg_nll/JSRT/README.md

@@ -71,6 +71,22 @@ pymic_run train config/unet_gce.cfg
 pymic_run test config/unet_gce.cfg
 ```
 
+### CLSLSR
+The CLSLSR method estimates errors in the original noisy label and obtains pixel-level weight maps based on an intial model, and then uses the weight maps to suppress noises in  a standard supervised learning procedure. Assume that the initial model is the baseline method, run the following command to obtain the weight maps:
+
+```bash
+python clslsr_get_condience config/unet_ce.cfg
+```
+
+The weight maps will be saved in `$root_dir/slsr_conf`. Then train the new model and do inference by:
+
+```bash
+pymic_run train config/unet_clslsr.cfg
+pymic_run test config/unet_clslsr.cfg
+```
+
+Note that the weight maps for training images are specified in the configuration file `train_csv = config/data/jsrt_train_mix_clslsr.csv`.
+
 ### Co-Teaching
 The configuration file for Co-Teaching is `config/unet2d_cot.cfg`. The corresponding setting is:
 
@@ -128,7 +144,7 @@ pymic_run test config/unet_dast.cfg
 Use `pymic_eval_seg config/evaluation.cfg` for quantitative evaluation of the segmentation results. You need to edit `config/evaluation.cfg` first, for example:
 
 ```bash
-metric = dice
+metric_list = [dice, assd]
 label_list = [255]
 organ_name = lung
 

seg_nll/JSRT/clslsr_get_confidence.py

@@ -0,0 +1,16 @@
+# -*- coding: utf-8 -*-
+from __future__ import print_function, division
+import sys
+from pymic.net_run_nll.nll_clslsr import get_confidence_map 
+
+
+if __name__ == "__main__":
+    """
+    The main function to get the confidence map during inference.
+    """
+    if(len(sys.argv) < 2):
+        print('Number of arguments should be 2. e.g.')
+        print('   python nll_clslsr.py config.cfg')
+        exit()
+    cfg_file = str(sys.argv[1])
+    get_confidence_map(cfg_file)