Prediction of the Ki-67 proliferation index in lung adenocarcinoma using an interpretable CT-based deep learning radiomics model: a two-center study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Foshan Hospital of Traditional Chinese Medicine, 6 Qinren Road, Chancheng District, Foshan, Guangdong Province, 528000, China.
- The Eighth Clinical Medical College of Guangzhou University of Chinese Medicine, Foshan , 528000, China.
- Department of Radiology, The First People's Hospital of Foshan, North Lingnan Avenue, Chancheng District, Foshan, Guangdong Province, 528000, China.
- Department of Radiology, The First People's Hospital of Foshan, North Lingnan Avenue, Chancheng District, Foshan, Guangdong Province, 528000, China. [email protected].
- Department of Radiology, Foshan Hospital of Traditional Chinese Medicine, 6 Qinren Road, Chancheng District, Foshan, Guangdong Province, 528000, China. [email protected].
- The Eighth Clinical Medical College of Guangzhou University of Chinese Medicine, Foshan , 528000, China. [email protected].
Abstract
To develop an interpretable computed tomography (CT)-based deep learning radiomics (DLR) model that integrates deep learning (DL) and radiomics features to predict the preoperative Ki-67 proliferation index in lung adenocarcinoma. A retrospective investigation involved 136 patients with lung adenocarcinoma who had undergone surgical resection between 2020 and 2024. Ki-67 expression was categorized as negative when ≤ 5%, and positive when > 5%. The training set comprised 91 patients, and the external test set contained 45 patients. Preoperative CT images were analyzed to derive 794 radiomics features, alongside 120 DL features obtained from the ResNet50 model. Model development was carried out by integrating DLR features through the application of ten distinct machine learning classifiers. The model’s diagnostic performance was evaluated based on the area under the curve (AUC). The interpretability was further examined using the Shapley additive explanations (SHAP) technique. Among the classifiers evaluated, the logistic regression (LR) model that integrates both DL and radiomics features exhibited the superior predictive performance, achieving an AUC of 0.904 in the training set, and the external test set AUC was 0.865, respectively. This result surpassed both traditional radiomics-based models (AUCs of 0.847 and 0.821 were observed for the training set and external test set, respectively) and DL models alone (AUCs of 0.802 and 0.733 were observed for the training set and external test set, respectively). Additionally, SHAP analysis demonstrated the two radiomics features with the greatest influence on model predictions. The CT-based DLR model demonstrates strong predictive capabilities for distinguishing Ki-67 expression status in lung adenocarcinoma, indicating its capacity as an interpretable and robust tool for preoperative evaluation. The online version contains supplementary material available at 10.1186/s12890-025-03983-5.