Application of 3D nnU-Net Model Based on Dual-Layer Detector CT-Guided Labeling in Non-Contrast CT Esophageal Segmentation: A Comparative Study with the Total Segmentator Model

Objective: To explore a 3D nnU Net model based on dual layer CT (DLCT) guided annotation for non contrast CT (NCCT) esophageal tissue with low contrast, difficult segmentation, and limited performance of general segmentation models, in order to improve the accuracy and generalization ability of NCCT esophageal segmentation and compare its performance with the Total Segmentor general model. Method: A retrospective study was conducted on 300 patients who underwent chest DLCT examination at the Fourth Affiliated Hospital of Nanjing Medical University. Virtual non contrast (VNC) and enhanced images were fused to achieve manual segmentation of the esophageal gold standard, and a 3D nnU Net model was trained using VNC. True non contrast (TNC) was used as the internal validation set; Additionally, 100 cases of NCCT from Southeast University Affiliated Zhongda Hospital and 48 cases of NCCT from the National Bioinformatics Center were included as two external test sets. Compare the segmentation performance of the two models using Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), consistency analysis, and Likert 5-point subjective scoring. Results: In the internal validation set and two external test sets, the DSC of our model was 0.756-0.798 and the HD95 was 4.00-4.10mm, both significantly better than the Total Segmentor model (P<0.001); Its consistency with the gold standard (Rspearman=0.821-0.887) and subjective score (5.00-5.00) were also significantly higher than the control model. Conclusion: The 3D nnU Net model based on DLCT guided annotation has better segmentation performance and generalization ability in esophageal segmentation, providing a new technical strategy for NCCT automated segmentation. Model training and source code:Brief code for training a CT-based deep learning segmentation model of thyroid and Papillary thyroid carcinoma using the nnU-Net framework. For more information on running inference with nnU-Net, click here: (https://github.com/MIC-DKFZ/nnUNet)。#0: activate virtual environmentconda create -n nnUNet python=3.10 -yconda activate nnUNetpip install nnunetv2pip install medpy#1: set environment variablesset nnUNet_raw_data_base="path\to\nnUNet_raw_data_base"set nnUNet_preprocessed="path\to\nnUNet_preprocessed"set RESULTS_FOLDER="path\to\nnUNet_trained_models"#2: Task11 (CT), Divide the data set"path\\Scripts\\nnUNetv2_convert_MSD_dataset.exe" -i "D:\nnunet\nnUNet\nnUNet_raw\Task11_example"#3: Preprocessing"path\\Scripts\\nnUNetv2_plan_and_preprocess.exe" -d 11 --verify_dataset_integrity#4: Training 3D fullres models call "path\Scripts\nnUNetv2_train.exe" 11 3d_fullres %%i -tr nnUNetTrainer -device cuda)#5: Test set reasoning 3D fullres"path\\Scripts\\nnUNetv2_predict.exe" -i D:\nnunet\nnUNet\nnUNet_raw\Dataset011_example\imagesTs -o output -d 11 -c 3d_fullres -f all#6: Compute the test set DSC"path\\Scripts\\nnUNetv2_evaluate_folder.exe" -djfile output\dataset.json -pfile output\plans.json D:\nnunet\nnUNet\nnUNet_raw\Task11_example\labelsTs output#7: Computing Test Set HD95"path\\python.exe" calculate_hd95_multilabel.py --pred_folder output --gt_folder D:\nnunet\nnUNet\nnUNet_raw\Task11_example\labelsTs --labels 1,2 --output hd95_multilabel_results.json

研究目标：针对低对比度、分割难度大且通用分割模型性能受限的非增强CT（Non-Contrast CT, NCCT）食管组织，探索基于双层CT（Dual Layer CT, DLCT）标注引导的3D nnU-Net模型，以提升NCCT食管分割的精度与泛化能力，并将其性能与通用分割模型Total Segmentor进行对比。

研究方法：本研究为回顾性研究，纳入南京医科大学第四附属医院行胸部双层CT检查的300例患者。通过融合虚拟非增强（Virtual Non-Contrast, VNC）与增强图像完成食管金标准的手动分割，并以VNC数据训练3D nnU-Net模型；以真实非增强CT（True Non-Contrast, TNC）作为内部验证集。此外纳入东南大学附属中大医院的100例NCCT病例与国家生物信息中心的48例NCCT病例作为两个外部测试集。采用戴斯相似性系数（Dice Similarity Coefficient, DSC）、95%豪斯多夫距离（95% Hausdorff Distance, HD95）、一致性分析及李克特5分量表（Likert 5-point scale）主观评分对两种模型的分割性能进行比较。

研究结果：在内部验证集与两个外部测试集中，本模型的DSC值为0.756~0.798，HD95值为4.00~4.10mm，两项指标均显著优于Total Segmentor模型（P<0.001）；其与金标准的一致性（斯皮尔曼相关系数Spearman Correlation Coefficient, Rspearman=0.821~0.887）及主观评分（5.00~5.00）也显著高于对照模型。

研究结论：基于DLCT标注引导的3D nnU-Net模型在食管分割任务中具备更优异的分割性能与泛化能力，为NCCT食管自动分割提供了全新的技术策略。

模型训练与源代码：本代码片段用于基于nnU-Net框架训练针对甲状腺及甲状腺乳头状癌的CT深度学习分割模型。如需了解nnU-Net推理运行的更多细节，请访问：https://github.com/MIC-DKFZ/nnUNet。

#0：激活虚拟环境
conda create -n nnUNet python=3.10 -y
conda activate nnUNet
pip install nnunetv2
pip install medpy

#1：设置环境变量
set nnUNet_raw_data_base="路径 o
nUNet_raw_data_base"
set nnUNet_preprocessed="路径 o
nUNet_preprocessed"
set RESULTS_FOLDER="路径 o
nUNet_trained_models"

#2：任务11（CT）：数据集划分
"路径Scripts
nUNetv2_convert_MSD_dataset.exe" -i "D:
nunet
nUNet
nUNet_rawTask11_example"

#3：数据预处理
"路径Scripts
nUNetv2_plan_and_preprocess.exe" -d 11 --verify_dataset_integrity

#4：训练3D全分辨率模型
调用 "路径Scripts
nUNetv2_train.exe" 11 3d_fullres %%i -tr nnUNetTrainer -device cuda

#5：3D全分辨率模型测试集推理
"路径Scripts
nUNetv2_predict.exe" -i D:
nunet
nUNet
nUNet_rawDataset011_exampleimagesTs -o output -d 11 -c 3d_fullres -f all

#6：计算测试集DSC值
"路径Scripts
nUNetv2_evaluate_folder.exe" -djfile outputdataset.json -pfile outputplans.json D:
nunet
nUNet
nUNet_rawTask11_examplelabelsTs output

#7：计算测试集HD95值
"路径python.exe" calculate_hd95_multilabel.py --pred_folder output --gt_folder D:
nunet
nUNet
nUNet_rawTask11_examplelabelsTs --labels 1,2 --output hd95_multilabel_results.json