Prediction of lymph node metastasis in invasive lung adenocarcinoma based on radiomics of the primary lesion, peritumoral region, and tumor habitat: A single-center retrospective study
10.7507/1007-4848.202503027
- VernacularTitle:基于原发灶、瘤周及生境影像组学预测浸润性肺腺癌淋巴结转移的单中心回顾性研究
- Author:
Hongchang WANG
1
;
Yan GU
1
;
Wenhao ZHANG
2
;
Guang MU
1
;
Wentao XUE
1
;
Mengen WANG
1
;
Chenghao FU
1
;
Liang CHEN
1
;
Mei YUAN
3
;
Jun WANG
1
Author Information
1. Department of Thoracic Surgery, Jiangsu Province Hospital (the First Affiliated Hospital with Nanjing Medical University), Nanjing, 210029, P. R. China
2. Department of Thoracic Surgery, Taihe Hospital, Shiyan, 442012, Hubei, P. R. China
3. Department of Radiology, Jiangsu Province Hospital (the First Affiliated Hospital with Nanjing Medical University), Nanjing, 210029, P. R. China
- Publication Type:Journal Article
- Keywords:
Invasive lung adenocarcinoma;
radiomics;
stacking model;
lymph node metastasis;
machine learning
- From:
Chinese Journal of Clinical Thoracic and Cardiovascular Surgery
2025;32(08):1079-1085
- CountryChina
- Language:Chinese
-
Abstract:
Objective To predict the lymph node metastasis status of patients with invasive pulmonary adenocarcinoma by constructing machine learning models based on primary tumor radiomics, peritumoral radiomics, and habitat radiomics, and to evaluate the predictive performance and generalization ability of different imaging features. Methods A retrospective analysis was performed on the clinical data of 1 263 patients with invasive pulmonary adenocarcinoma who underwent surgery at the Department of Thoracic Surgery, Jiangsu Province Hospital, from 2016 to 2019. Habitat regions were delineated by applying K-means clustering (average cluster number of 2) to the grayscale values of CT images. The peritumoral region was defined as a uniformly expanded area of 3 mm around the primary tumor. The primary tumor region was automatically segmented using V-net combined with manual correction and annotation. Subsequently, radiomics features were extracted based on these regions, and stacked machine learning models were constructed. Model performance was evaluated on the training, testing, and internal validation sets using the area under the receiver operating characteristic curve (AUC), F1 score, recall, and precision. Results After excluding patients who did not meet the screening criteria, a total of 651 patients were included. The training set consisted of 468 patients (181 males, 287 females) with an average age of (58.39±11.23) years, ranging from 29 to 78 years, the testing set included 140 patients (56 males, 84 females) with an average age of (58.81±10.70) years, ranging from 34 to 82 years, and the internal validation set comprised 43 patients (14 males, 29 females) with an average age of (60.16±10.68) years, ranging from 29 to 78 years. Although the habitat radiomics model did not show the optimal performance in the training set, it exhibited superior performance in the internal validation set, with an AUC of 0.952 [95%CI (0.87, 1.00)], an F1 score of 84.62%, and a precision-recall AUC of 0.892, outperforming the models based on the primary tumor and peritumoral regions. Conclusion The model constructed based on habitat radiomics demonstrated superior performance in the internal validation set, suggesting its potential for better generalization ability and clinical application in predicting lymph node metastasis status in pulmonary adenocarcinoma.
