1st place solution in MICCAI 2020 TN-SCUI challenge

This is the source code of the 1st place solution for segmentation task (with IoU 82.54%) in 2020 TN-SCUI challenge.

[Challenge leaderboard🏆] [中文版readme👁]

Fold1 test results of DDTI dataset in colab

Pipeline of our solution

We use a simple cascaded framework for segmenting nodules, it can be easily extended to other single-target segmentation tasks.

Our pipeline is shown in the figure below.

In summary, what we do in our solution are:

• preprocessing to remove irrelevant regions
• using a cascaded framework
• 5-fold cross-validation (CV) strategy with a balanced nodule size and category distribution
• using test time augmentation (TTA)
• model ensembling: since we trained two networks separately in 5-fold CV , we combined any one first network and one second network as a pair, and finally we got 25 pairs (or inference results). We use step4_Merge.py to merge 25 inference results into a final ensemble result by pixel-wised voting

Segmentation results on 2020 TN-SCUI training dataset and DDTI dataset

Reproducing the experiment results

You can follow the step-by-step tutorial below, or just download the pretrained weights and run test_fold1.py to reproduce our results. (note that we only provide the weights of 1st fold in 5-fold CV).

• The pretrained weights (DeeplabV3+ with efficientnet-B6 encoder) at stage Ⅰ and Ⅱ trained on TN-SCUI and DDTI dataset of fold1 are available at [GoogleDrive] [BaiduWP, password:qxds]
• The predicted segmentation masks of TN-SCUI final testing dataset (final submission with IoU 82.45%) are available at [GoogleDrive] [BaiduWP, password:qxds]

• We have processed the DDTI dataset into the format of 2020 TN-SCUI dataset，you can download it from [GoogleDrive] [BaiduWP, password:qxds]

  Fold1 test results of DDTI dataset in colab

  Statement

  We consulted the official organizer of 2020 TN-SCUI via email, and the official reply is that using of external dataset is not allowed in the 2020 TN-SCUI challenge, and participants are not allowed to provide the challenge data anywhere. Therefore, we hereby declare that we did not use any external datasets (such as the DDTI dataset) in the competition, and will not provide TNSCUI data anywhere. The DDTI dataset processed by us is only used as demo data for this code, so that everyone can run the entire scripts with a ready-made data.

How to run this code on the TN-SCUI dataset or your own custom dataset?

This code can be easily performed on single-target segmentation tasks. Here, we split the whole process into 5 steps so that you can easily replicate our results or perform the whole pipeline on your private custom dataset.

• step0, preparation of environment
• step1, run the script step1_preprocessing.m to perform the preprocessing
• step2, run the script step2_TrainAndValidate.py to train and validate the CNN model
• step3, run the script step3_TestOrInference.py to testing model on original unprocessed images
• step4 (optional), run the script step4_Merge.py to get ensemble predicted result

You should prepare your data in the format of 2020 TN-SCUI dataset, this is very simple, you only need to prepare:

• two folders store PNG format images and masks respectively. Note that the file names (ID) of the images and masks should be integers and correspond to each other
• a csv file that stores the image names (ID) and category. If there is only a segmentation task without a classification task, you can just falsify the category. For example, half of the data randomly is positive , and the other half is negative

Since the sponsor of TNSCUI challenge does not allow the dissemination of the 2020 TN-SCUI dataset, we give another thyroid ultrasound dataset demo for your reference. The demo dataset was derived from the DDTI open access dataset, and we have processed the dataset into the format of 2020 TN-SCUI dataset，you can download it from [GoogleDrive] [BaiduWP, password:qxds].

Step0 preparation of environment

We have tested our code in following environment：

• segmentation_models_pytorch == 0.1.0 (with install command pip install segmentation-models-pytorch)
• ttach == 0.0.3
• torch >＝1.0.0
• torchvision
• imgaug

For installing segmentation_models_pytorch (smp), there are two commands：

pip install segmentation-models-pytorch

Use the above command, you can install the released version that can run with torch <= 1.1.0, but without some decoders such as deeplabV3, deeplabV3+ and PAN.

pip install git+https://github.com/qubvel/segmentation_models.pytorch

Use the above command, you will install the latest version which includes decoders such as deeplabV3, deeplabV3+ and PAN, but requiring torch >= 1.2.0.

We modified the original code of smp, to make it can be run with torch <= 1.1.0, and also have the latest version of the smp library encoder such as deeplabV3, deeplabV3+ and PAN. The modified smp is in the filefolder segmentation_models_pytorch_4TorchLessThan120, here are two ways to use it:

1. install smp with pip install segmentation-models-pytorch firstly, then just copy the segmentation_models_pytorch_4TorchLessThan120 into your project folder to use it
2. copy the segmentation_models_pytorch_4TorchLessThan120 into your project folder, then install these libs: torchvision>=0.3.0, pretrainedmodels==0.7.4, efficientnet-pytorch>=0.5.1

Step1 preprocessing

👀This step is optional, if you donot wanna preprocess image, just prepare your data as the demo format and skip to the Step2"training and validatation".

In step1, you should run the script step1_preprocessing.m in MATLAB to perform the preprocessing. For both TN-SCUI and DDTI dataset, there are some irrelevant regions to be removed. Because we used a cross-validation with a nodule size and category balance strategy, we should get the size of nodules for cross-validation.

After the code runs, you will get two folders and a csv file:

• the two filefolders are named stage1 and stage2, the data in folder stage1 is used to train the first network, which contains the preprocessed image with no irrelevant regions; and stage2 is used to train the second network, which contains the image in the expanded ROI
• the csv file named train.csv with below dataformat, the size of a nodule is the number of pixels of the nodule after unifying preprocessed image to 256×256 pixels

The example of preprocessed DDTI dataset was given here [GoogleDrive] [BaiduWP, password:qxds]

Step2 training and validatation

For training phase：

In step2, you should run the script step2_TrainAndValidate.py to train the first or second network in the cascaded framework (you need to train the two networks separately), that is, the training process (such as hyperparameters) is the same for both first and second network except for the different training datasets used. With the csv file train.csv, you can directly perform K-fold cross validation (default is 5-fold), and the script uses a fixed random seed to ensure that the K-fold cv of each experiment is repeatable (please note that the same random seed may cause different K-fold under different versions of scikit-learn).

For example, you can train the first fold of the first network in two ways:

1. modify (set) following parameters in the script step2TrainAndValidate.py and run the script.

   """ just set your param as the default value """
   # hyperparameters
   parser.add_argument('--image_size', type=int, default=256)  # recommended: 256 for 1st(stageⅠ) network, 512 for 2nd(stage Ⅱ) network👍
   parser.add_argument('--batch_size', type=int, default=3)
   parser.add_argument('--batch_size_test', type=int, default=30)
   # custom parameters
   parser.add_argument('--Task_name', type=str, default='dpv3plus_stage1_', help='DIR name,Task name')
   parser.add_argument('--csv_file', type=str, default='./DDTI/2_preprocessed_data/train.csv')
   parser.add_argument('--filepath_img', type=str, default='./DDTI/2_preprocessed_data/stage1/p_image')
   parser.add_argument('--filepath_mask', type=str, default='./DDTI/2_preprocessed_data/stage1/p_mask')
   parser.add_argument('--fold_K', type=int, default=5, help='number of cross-validation folds')
   parser.add_argument('--fold_idx', type=int, default=1)

2. use the following command at the terminal to transmit parameters and run the script.

   python step2_TrainAndValidate.py \
   --batch_size=3 \
   --Task_name='dpv3plus_stage1_'\
   --csv_file='./DDTI/2_preprocessed_data/train.csv' \
   --filepath_img='./DDTI/2_preprocessed_data/stage1/p_image' \
   --filepath_mask='./DDTI/2_preprocessed_data/stage1/p_mask' \
   --fold_idx=1

Something else in training process:

• when the GPU is geforce 1080ti, the training batchsize can be set as 10 for first network and 3 for second network, and testing batchsize can be set as 30 for both network
• the weights with the highest IoU on the validation set will be saved
• by specifying the parameter fold_idx from 1 to K and running repeatedly, you can complete the training of all K folds. The results will be saved in the result folder under the project. You can view the training log in real time through tensorboardX
• for training the second network, you only need to change the path of the training data (filepath_img and filepath_mask) and Task_name
• we do not validation set during training, actually only use the training set and the test set, and copy the test set to the validation set. If you need to use a validation set, you can specify the parameter validate_flag as True in the script step2TrainAndValidate.py
• if you do not want to perform cross-validation, or have already prepared a train and validate and test set, just modify the code in the funtion main() with following code:

  train_list = get_filelist_frompath(train_img_folder,'PNG') 
  train_list_GT = [train_mask_folder+sep+i.split(sep)[-1] for i in train_list]
  test_list = get_filelist_frompath(test_img_folder,'PNG') 
  test_list_GT = [test_mask_folder+sep+i.split(sep)[-1] for i in test_list]

For validation and testing phase

In fact, there is no need to re-run the script, because the test has been performed during the training process, and the detailed information of the result is recorded in the record.xlsx or record.txt. But you can also confirm the final training effect by running the script step2TrainAndValidate.py after the training by setting the parameter mode as test.

It should be noted that the dataset validated in the training phase is preprocessed. Therefore, if you want to perform test based on the original unprocessed images, please refer to step3.

Step3 testing (or inference or predicting)

In step3, you should run the script step3_TestOrInference.py to perform inference based on the original unprocessed image. By using the script step3_TestOrInference.py, you can:

• perform test of cross-validation (based on the original unprocessed images)
• inference on new data

Here are some parameters should be noticed (Especially c1_size and c2_size):

fold_flag = False # False for inference on new data, and True for cross-validation
mask_path = None  # None for inference on new data
save_path = None  # If you don't want to save the inference result, set it to None
c1_size = 256  # The input image size for training the first network ❗
c1_tta = False # Whether to use TTA for the first network
orimg=False    # False for no preprocessing to remove irrelevant areas
use_c2_flag = False # Whether to use the second network
c2_size = 512 # The input image size for training the second network ❗
c2_tta = False # Whether to use TTA for the second network

When you get multiple results after inference, you can use the script step4_Merge.py to integrate all the results. We use a pixel-wised voting method, and recommend voting on odd numbers of results to avoid fluctuations in making decision.

Acknowledgement

• Thanks to the organizers of the 2020 TN-SCUI challenge
• Thanks to the open access DDTI dataset for providing ultrasound data and annotation of thyroid nodules
• Thanks to qubvel, the author of smg and ttach, all network and TTA used in this code come from his implement