Transfer Mapping for Clinically Relevant Lymph Nodes in The Mediastinal Area of CT data
Hanxiao Zhang, Minghui Zhang, Xin You, Zhebing Lin, Yi Zhang, Liu Liu, Xinghua Cheng, Yun Gu and Guang-Zhong Yang
Built upon nnUNet, this repository provides the solution of Team IMR for MICCAI 2023 LNQ Challenge. This repository contains the training method and inference code for diseased and mediastinal lymph node segmentation using contrast-enhanced CT scans.
In that LNQ training data are partially annotated (i.e. one node out of five) and it is suboptimal to use straight off the shelf for fully supervised pixel-wise segmentation, this solution leverages two additional data with full mediastinal lymph node annotations in the training process. Both additional data are publicly available. Totally 102 CT volumes (87 CTs + 15 CTs) with full annotations were employed.
TCIA data with refined annotations
Original TCIA data consists of 90 contrast-enhanced CTs of the mediastinum with lymph node position labels marked by radiologists at the National Institutes of Health (Roth et al. 2014). Based on these 90 CTs, Self et al. 2015 independently produced lymph node mask annotations examined by a board-certified radiologist. However, these annotations are often sparse and not available with a complete set of mediastinal lymph nodes (e.g., some small but visible lymph nodes were left unsegmented). Bouget et al. 2023 took these available annotations as a starting point and manually refined segmentations for all mediastinal lymph nodes in 89 CTs (mediastinal case 43 was removed by Bouget et al. 2023 due to its incomplete CT volume). We additionally excluded two CTs (mediastinal case 06 and case 80) because of the absence of full annotations after strict examination, thus leaving eligible a final set of 87 CTs.
St. Olavs Hospital data
Original St. Olavs Hospital data consists of 17 CTs, among which 15 contrast-enhanced CTs were first assigned their manual annotations of lymph nodes and fifteen different anatomical structures in the mediastinal area (Bouget et al. 2019). These lymph node annotations of the 15 CTs were further refined and proofed by an expert thoracic radiologist, as a benchmark dataset of Bouget et al. 2023.
Install nnUNet as below and meet the requirements of nnUNet. For more details please refer to https://github.com/MIC-DKFZ/nnUNet/tree/nnunetv1
git clone https://github.com/MIC-DKFZ/nnUNet.git
cd nnUNet
git checkout nnunetv1
pip install -e .
- Resampling (0.8mm x 0.8mm x 0.8mm)
- Lung segmentation
- Initial VOI boundary extraction based on lung mask
- Cropping VOI for lung airway segmentation
- Removing airway maps in lung mask regions
- Secondary VOI boundary extraction based on bronchus
- Cropping input VOI based on two VOI bounding boxes
- Developing a nnUNet model (3d_fullres configuration) based on two additional data with full annotations (102 cases)
- Training nnUNet based on one fold of the 5-fold-cross-validation for 1000 epochs
- Resampling and restoring to original CT volume
- Removing individual components whose volume are less than the volume of a sphere with a radius of 5 mm
- Generating pseudo labels for LNQ training data
- Combining the LNQ pseudo labels and original labels to generate the new labels
- The total cohort includes the full-annotation data and LNQ data with pseudo label transfer
- Finetune the former model parameters using this total cohort for 300 epochs