1st place solution in CCF BDCI 2021 ULSEG challenge

Overview

1st place solution in CCF BDCI 2021 ULSEG challenge

This is the source code of the 1st place solution for ultrasound image angioma segmentation task (with Dice 90.32%) in 2021 CCF BDCI challenge.

[Challenge leaderboard 🏆 ]

Pipeline of our solution

Our solution includes data pre-processing, network training, ensabmle inference and post-processing.

Data pre-processing

To improve our performance on the leaderboard, 5-fold cross validation is used to evaluate the performance of our proposed method. In our opinion, it is necessary to keep the size distribution of tumor in the training and validation sets. We calculate the tumor area for each image and categorize the tumor size into classes: 1) less than 3200 pixels, 2) less than 7200 pixels and greater than 3200 pixels, and 3) greater than 7200 pixels. These two thresholds, 3200 pixels and 7200 pixels, are close to the tertiles. We divide images in each size grade group into 5 folds and combined different grades of single fold into new single fold. This strategy ensured that final 5 folds had similar size distribution.

Network training

Due to the small size of the training set, for this competition, we chose a lightweight network structure: Linknet with efficientnet-B6 encoder. Following methods are performed in data augmentation (DA): 1) horizontal flipping, 2) vertical flipping, 3) random cropping, 4) random affine transformation, 5) random scaling, 6) random translation, 7) random rotation, and 8) random shearing transformation. In addition, one of the following methods was randomly selected for enhanced data augmentation (EDA): 1) sharpening, 2) local distortion, 3) adjustment of contrast, 4) blurring (Gaussian, mean, median), 5) addition of Gaussian noise, and 6) erasing.

Ensabmle inference

We ensamble five models (five folds) and do test time augmentation (TTA) for each model. TTA generally improves the generalization ability of the segmentation model. In our framework, the TTA includes vertical flipping, horizontal flipping, and rotation of 180 degrees for the segmentation task.

Post-processing

We post-processe the obtained binary mask by removing small isolated points (RSIP) and edge median filtering (EMF) . The edge part of our predicted tumor is not smooth enough, which is not quite in line with the manual annotation of the physician, so we adopt a small trick, i.e., we do a median filtering specifically for the edge part, and the experimental results show that this can improve the accuracy of tumor segmentation.

Segmentation results on 2021 CCF BDCI dataset

We test our method on 2021 CCD BDCI dataset (215 for training and 107 for testing). The segmentation results of 5-fold CV based on "Linknet with efficientnet-B6 encoder" are as following:

fold Linknet Unet Att-Unet DeeplabV3+ Efficient-b5 Efficient-b6 Resnet-34 DA EDA TTA RSIP EMF Dice (%)
1 85.06
1 84.48
1 84.72
1 84.93
1 86.52
1 86.18
1 86.91
1 87.38
1 88.36
1 89.05
1 89.20
1 89.52
E 90.32

How to run this code?

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 preprocess.py to perform the preprocessing
  • step2, run the script train.py to train our model
  • step3, run the script inference.py to inference the test data.
  • step4, run the script postprocess.py to perform the preprocessing.

You should prepare your data in the format of 2021 CCF BDCI dataset, this is very simple, you only need to prepare: two folders store png format images and masks respectively. You can download them from [Homepage].

The complete file structure is as follows:

  |--- CCF-BDCI-2021-ULSEG-Rank1st
      |--- segmentation_models_pytorch_4TorchLessThan120
          |--- ...
          |--- ...
      |--- saved_model
          |--- pred
          |--- weights
      |--- best_model
          |--- best_model1.pth
          |--- ...
          |--- best_model5.pth
      |--- train_data
          |--- img
          |--- label
          |--- train.csv
      |--- test_data
          |--- img
          |--- predict
      |--- dataset.py
      |--- inference.py
      |--- losses.py
      |--- metrics.py
      |--- ploting.py
      |--- preprocess.py
      |--- postprocess.py
      |--- util.py
      |--- train.py
      |--- visualization.py
      |--- requirement.txt

Step0 preparation of environment

We have tested our code in following environment:

For installing these, run the following code:

pip install -r requirements.txt

Step1 preprocessing

In step1, you should run the script and train.csv can be generated under train_data fold:

python preprocess.py \
--image_path="./train_data/label" \
--csv_path="./train_data/train.csv"

Step2 training

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. Run the following code:

python train.py \
--input_channel=1 \
--output_class=1 \
--image_resolution=256 \
--epochs=100 \
--num_workers=2 \
--device=0 \
--batch_size=8 \
--backbone="efficientnet-b6" \
--network="Linknet" \
--initial_learning_rate=1e-7 \
--t_max=110 \
--folds=5 \
--k_th_fold=1 \
--fold_file_list="./train_data/train.csv" \
--train_dataset_path="./train_data/img" \
--train_gt_dataset_path="./train_data/label" \
--saved_model_path="./saved_model" \
--visualize_of_data_aug_path="./saved_model/pred" \
--weights_path="./saved_model/weights" \
--weights="./saved_model/weights/best_model.pth" 

By specifying the parameter k_th_fold from 1 to folds and running repeatedly, you can complete the training of all K folds. After each fold training, you need to copy the .pth file from the weights path to the best_model folder.

Step3 inference (test)

Before running the script, make sure that you have generated five models and saved them in the best_model folder. Run the following code:

python inference.py \
--input_channel=1 \
--output_class=1 \
--image_resolution=256 \
--device=0 \
--backbone="efficientnet-b6" \
--network="Linknet" \
--weights1="./saved_model/weights/best_model1.pth" \
--weights2="./saved_model/weights/best_model2.pth" \
--weights3="./saved_model/weights/best_model3.pth" \
--weights4="./saved_model/weights/best_model4.pth" \
--weights5="./saved_model/weights/best_model5.pth" \
--test_path="./test_data/img" \
--saved_path="./test_data/predict" 

The results of the model inference will be saved in the predict folder.

Step4 postprocess

Run the following code:

python postprocess.py \
--image_path="./test_data/predict" \
--threshood=50 \
--kernel=20 

Alternatively, if you want to observe the overlap between the predicted result and the original image, we also provide a visualization script visualization.py. Modify the image path in the code and run the script directly.

Acknowledgement

  • Thanks to the organizers of the 2021 CCF BDCI challenge.
  • Thanks to the 2020 MICCCAI TNSCUI TOP 1 for making the code public.
  • Thanks to qubvel, the author of smg and ttach, all network and TTA used in this code come from his implement.
You might also like...
The sixth place winning solution (6/220) in 2021 Gaofen Challenge.
The sixth place winning solution (6/220) in 2021 Gaofen Challenge.

SwinTransformer + OBBDet The sixth place winning solution (6/220) in the track of Fine-grained Object Recognition in High-Resolution Optical Images, 2

Codebase for the solution that won first place and was awarded the most human-like agent in the 2021 NeurIPS Competition MineRL BASALT Challenge.

KAIROS MineRL BASALT Codebase for the solution that won first place and was awarded the most human-like agent in the 2021 NeurIPS Competition MineRL B

QQ Browser 2021 AI Algorithm Competition Track 1 1st Place Program

QQ Browser 2021 AI Algorithm Competition Track 1 1st Place Program

4th place solution to datafactory challenge by Intermarché.

Solution to Datafactory challenge by Intermarché. 4th place solution to datafactory challenge by Intermarché. The objective of the challenge is to pre

Kaggle | 9th place (part of) solution for the Bristol-Myers Squibb – Molecular Translation challenge

Part of the 9th place solution for the Bristol-Myers Squibb – Molecular Translation challenge translating images containing chemical structures into I

Kaggle | 9th place single model solution for TGS Salt Identification Challenge

UNet for segmenting salt deposits from seismic images with PyTorch. General We, tugstugi and xuyuan, have participated in the Kaggle competition TGS S

10th place solution for Google Smartphone Decimeter Challenge at kaggle.
10th place solution for Google Smartphone Decimeter Challenge at kaggle.

Under refactoring 10th place solution for Google Smartphone Decimeter Challenge at kaggle. Google Smartphone Decimeter Challenge Global Navigation Sat

Kohei's 5th place solution for xview3 challenge

xview3-kohei-solution Usage This repository assumes that the given data set is stored in the following locations: $ ls data/input/xview3/*.csv data/in

4st place solution for the PBVS 2022 Multi-modal Aerial View Object Classification Challenge - Track 1 (SAR) at PBVS2022
4st place solution for the PBVS 2022 Multi-modal Aerial View Object Classification Challenge - Track 1 (SAR) at PBVS2022

A Two-Stage Shake-Shake Network for Long-tailed Recognition of SAR Aerial View Objects 4st place solution for the PBVS 2022 Multi-modal Aerial View Ob

Owner
Chenxu Peng
Data Science, Deep Learning
Chenxu Peng
The 1st place solution of track2 (Vehicle Re-Identification) in the NVIDIA AI City Challenge at CVPR 2021 Workshop.

AICITY2021_Track2_DMT The 1st place solution of track2 (Vehicle Re-Identification) in the NVIDIA AI City Challenge at CVPR 2021 Workshop. Introduction

Hao Luo 91 Dec 21, 2022
Code for 1st place solution in Sleep AI Challenge SNU Hospital

Sleep AI Challenge SNU Hospital 2021 Code for 1st place solution for Sleep AI Challenge (Note that the code is not fully organized) Refer to the notio

Saewon Yang 13 Jan 3, 2022
1st place solution to the Satellite Image Change Detection Challenge hosted by SenseTime

1st place solution to the Satellite Image Change Detection Challenge hosted by SenseTime

Lihe Yang 209 Jan 1, 2023
My 1st place solution at Kaggle Hotel-ID 2021

1st place solution at Kaggle Hotel-ID My 1st place solution at Kaggle Hotel-ID to Combat Human Trafficking 2021. https://www.kaggle.com/c/hotel-id-202

Kohei Ozaki 18 Aug 19, 2022
🏆 The 1st Place Submission to AICity Challenge 2021 Natural Language-Based Vehicle Retrieval Track (Alibaba-UTS submission)

AI City 2021: Connecting Language and Vision for Natural Language-Based Vehicle Retrieval ?? The 1st Place Submission to AICity Challenge 2021 Natural

null 82 Dec 29, 2022
Xview3 solution - XView3 challenge, 2nd place solution

Xview3, 2nd place solution https://iuu.xview.us/ test split aggregate score publ

Selim Seferbekov 24 Nov 23, 2022
1st Place Solution to ECCV-TAO-2020: Detect and Represent Any Object for Tracking

Instead, two models for appearance modeling are included, together with the open-source BAGS model and the full set of code for inference. With this code, you can achieve around mAP@23 with TAO test set (based on our estimation).

null 79 Oct 8, 2022
Waymo motion prediction challenge 2021: 3rd place solution

Waymo motion prediction challenge 2021: 3rd place solution ?? Technical report ??️ Presentation ?? Announcement ??Motion Prediction Channel Website ??

null 158 Jan 8, 2023
4th place solution for the SIGIR 2021 challenge.

SIGIR-2021 (Tinkoff.AI) How to start Download train and test data: https://sigir-ecom.github.io/data-task.html Place it under sigir-2021/data/. Run py

Tinkoff.AI 4 Jul 1, 2022
Meli Data Challenge 2021 - First Place Solution

My solution for the Meli Data Challenge 2021

Matias Moreyra 23 Mar 9, 2022