Solution of team HIASBUAA of FLARE2023 Challenge

Coarse to Fine Segmentation Method Enables Accurate and Efficient Segmentation of Organs and Tumor in Abdominal CT

Hui Meng, Haochen Zhao, Deqian Yang, Wangsong ping, Zhenpeng Li

this is the official code of Team hiasbuaa for FLARE23 Challenge

Overview of our work.

we use coarse to fine three stage method to segment organs and tumors.

we train a coarse model to make the focus of ct is on the abdomial area.
- coarse model was trained on Group6 219 CTs
we train a tumor model to segment tumors.
- using coarse model to inference the Group4 1497 CTs to make a dataset with full organ and tumor.
- tumor model was trained on the above dataset
we train a organ model to segment organs.
- organ model was trained on Group1 and Group2 4000 CTs ,which pseudo-label was obtained from flare committee (blackbean)

Dataset

we divide the data in to 6 groups

Preprocessing

A brief description of the preprocessing method

cropping: Before model training, the training CT scans are cropped along the z-axis direction based on ground truth or pseudo labels. Specifically, the indices of start slice and the end slice of region containing targets are first calculated based on labels. To reserve context information of segmentation targets, we reduce the index of the start slice by 10 and add the index of the end slice by 10 During model training, the cropped CT scans are further cropped based on non-zero region introduced by nnU-Net.
intensity normalization: We gather pixel values in the cropped CT scans and subsequently truncateall data to fall within [0.5, 99.5] of foreground voxel values .Following that,z-score normalization is applied.
Resampling method for anisotropic data: We perform image redirection to the desired orientation, followed by resampling all CT scans to match the median voxel spacing of the training dataset. Specifically, third-order spline interpolation is used for image resampling, and the nearest neighbor interpolation is employed for label resampling.

Environments and Installation

Ubuntu 20.04 LTS
CPU 13th Gen Intel(R) Core(TM) i7-13700KF 3.40 GHz
RAM 32GB;
GPU 1 NVIDIA RTX 4090 12G
CUDA version 12.1
python version 3.10

Training

cd nnunet/experiment_planning/
python nnUNet_convert_decathlon_task.py
python nnUNet_plan_and_preprocess.py
python run_training.py 3d_lowres nnUNetTrainerV2

Inference

we use three stage framework to inference the online 100 validation set. 4 pretrained weights was provided to get the results, It is also needed to put the weights to the path nnunet/inference/engine/. which is provided on the zenodo

python predict_final.py

Results

Our method achieves the following performance on FLARE23: Fast, Low-resource, and Accurate oRgan and Pan-cancer sEgmentation in Abdomen CT

Model name	Organs DICE	Tumor DICE	Organ NSD	Organ DSD
our method	91.59%	47.12%	95.74%	39.94

Reference

MACCAI FLARE2023 https://codalab.lisn.upsaclay.fr/competitions/12239

MACCAI FLARE2022 Team balackbean https://github.com/Ziyan-Huang/FLARE22

zhuji423 / MICCAI2023_Flare2023