AbdomenCT/3-WeaklySupervisedLearning

• Weakly Supervised Learning Baseline
• • Introduction
  • Training
  • Trained Models
  • Inference
  • Baseline Docker
  • CRF Postprocessing
  • Reference
  • Acknowledgements

Weakly Supervised Learning Baseline

Introduction

The baseline is built on the combination of 2D nnUNet [1] and fully connected Conditional Random Fields (CRF)[2], which is motivated by Gao et al.'s method in addressing the missing annotation problem [3].

Training

Run

nnUNet_train 2d nnUNetTrainerV2 TaskXXX_MYTASK all

Trained Models

• Dropbox
• Baidu Net Disk (PW:2021)

Inference

Run

nnUNet_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -t TASK_NAME_OR_ID -m 2d -disable_tta -f all --save_npz

Baseline Docker

• Subtask 1: 5% labels

Run

# pull docker image from hub.docker.com
docker pull gospelslave/weakly_suplearn_subtask1

# docker predict command
docker container run --gpus "device=1" --name weakly_suplearn_subtask1 --rm -v $PWD/TestImage/:/workspace/input/ -v $PWD/weakly_suplearn_subtask1_outputs/:/workspace/outputs/ gospelslave/weakly_suplearn_subtask1:latest /bin/bash -c "sh predict.sh"

• Subtask 2: 15% labels

Run

# pull docker image from hub.docker.com
docker pull gospelslave/weakly_suplearn_subtask2

# docker predict command
docker container run --gpus "device=1" --name weakly_suplearn_subtask2 --rm -v $PWD/TestImage/:/workspace/input/ -v $PWD/weakly_suplearn_subtask2_outputs/:/workspace/outputs/ gospelslave/weakly_suplearn_subtask2:latest /bin/bash -c "sh predict.sh"

• Subtask 3: 30% labels

Run

# pull docker image from hub.docker.com
docker pull gospelslave/weakly_suplearn_subtask3

# docker predict command
docker container run --gpus "device=1" --name weakly_suplearn_subtask3 --rm -v $PWD/TestImage/:/workspace/input/ -v $PWD/weakly_suplearn_subtask3_outputs/:/workspace/outputs/ gospelslave/weakly_suplearn_subtask3:latest /bin/bash -c "sh predict.sh"

CRF Postprocessing

Set input and output path and run

python AbdominalOrganCRFSeg.py

Reference

[1] F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, "nnU-Net: a self-confifiguring method for deep learning" based biomedical image segmentation,” Nature Methods, vol. 18, no. 2, pp. 203–211, 2021.

[2] P. Krahenbuhl and V. Koltun, “Efficient inference in fully connected crfs with gaussian edge potentials,” vol. 24, pp. 109–117, 2011.

[3] M. Gao, Z. Xu, L. Lu, A. Wu, I. Nogues, R. M. Summers, and D. J. Mollura, “Segmentation label propagation using deep convolutional neural networks and dense conditional random field,” in 2016 IEEE 13th International Symposium on Biomedical Imaging, pp. 1265–1268, 2016.

Acknowledgements

We highly appreciate the out-of-the-box implementation of nnU-Net (by Dr. Fabian Isensee) and CRF (by Dr. Guotai Wang).