MRI-derived habitat radiomics integrated with transformer deep learning for preoperative prediction of high CD9 expression: a novel prognostic marker in adult-type diffuse gliomas.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
- Department of CT/MR, Xingtai People's Hospital, Xingtai, China.
- Department of Neurology, Xingtai People's Hospital, Xingtai, China.
- Department of Neuro-oncology, Cancer Center, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
Abstract
High CD9 expression is significantly associated with shorter overall survival (OS) in patients with adult-type diffuse gliomas (ADGs). Concurrently, drug screening analyses have identified emetine as a potential targeted therapeutic agent for gliomas exhibiting high CD9 expression, indicating that CD9 may serve as a novel biomarker with the potential to transform glioma diagnosis and treatment. This study aimed to develop an artificial intelligence prediction model based on preoperative magnetic resonance imaging (MRI) images to facilitate the early identification of high CD9 expression in patients with ADGs. We retrospectively collected CD9 expression data from 262 patients with ADGs in The Cancer Genome Atlas (TCGA) and Chinese Glioma Genome Atlas (CGGA) cohorts. Using preoperative MRI sequences [contrast-enhanced T1-weighted imaging (CET1) and T2-weighted imaging (T2WI)], Habitat features were extracted, using K-means clustering algorithm, together with radiomic features. Additionally, multi-channel 2.5-dimensional deep learning (2.5D_DL) data were constructed, based on neural network architectures, including ResNet101, DenseNet121, and SimpleViT, and DL features were extracted. Seven classifiers were utilized to construct individual Radiomics, Habitat, and 2.5D_DL predictive models. Subsequently, a Stacking model was developed via an ensemble fusion strategy to predict high CD9 expression. Model performance was evaluated using receiver operating characteristic (ROC) curves, and clinical efficacy was assessed by decision curve analysis (DCA). Shapley Additive Explanations (SHAP) analysis was performed to explore the interpretation of the models, enhancing the transparency of the machine learning predictions. In addition, Patients were stratified into high- and low-risk groups based on Stacking model scores, and survival differences were compared using the Kaplan-Meier analysis and the log-rank test. The Habitat model exhibited superior ROC curve [area under the ROC (AUC)] values than the Radiomics model. Among the multi-channel 2.5D_DL models, the SimpleViT model achieved slightly better performance than the convolutional neural network (CNN)-based models (ResnNet101 and DenseNet121). The final Stacking model, which integrated the outputs from Habitat, Radiomics, and 2.5D_DL models exhibited superior AUC values. The AUC reached 0.999, 0.822, and 0.728 in the training, validation, and test sets, respectively. DeLong testing confirmed the statistically significant differences in the AUC values of each model in the training set, but no statistically significant differences were observed among the other sets. In addition, enhanced net clinical benefit across a more extensive range of risk thresholds for this integrated model was verified by DCA. Prognostic stratification by Stacking model revealed significant differences in survival outcomes between high-risk and low-risk groups. The preoperative MRI-based Stacking model effectively identifies patients with high CD9 expression among ADGs and stratifies patient risk, providing support for clinical therapeutic decision-making and individualized prognostic evaluation. Prospective multicenter data are still required to validate the prognostic performance and generalizability of this model.