Multi-dimensional interpretable deep learning-radiomics based on intra-tumoral and spatial habitat for preoperative prediction of thymic epithelial tumours risk categorisation.

BACKGROUND AND PURPOSE

This study aims to develop and compare combined models based on enhanced CT-based radiomics, multi-dimensional deep learning, clinical-conventional imaging and spatial habitat analysis to achieve accurate prediction of thymoma risk classification.

MATERIALS AND METHODS

205 consecutive patients with thymoma confirmed by surgical pathology were recruited from three medical centers. Venous phase enhanced CT images were used to delineate the tumor, and radiomics, 2D and 3D deep learning models based on the whole tumor were established and feature extraction was performed. The tumors were divided into different sub-regions by K-means clustering method and the corresponding features were obtained. The clinical-conventional imaging data of the patients were collected and evaluated, and the univariate and multivariate analysis were used for screening. The above types of features were fused with each other to construct a variety of combined models. Quantitative indicators such as area under the receiver operating characteristic (ROC) curve (AUC) were calculated to evaluate the performance of the model.

RESULTS

The AUC of RDLCSM developed based on LightGBM classifier was 0.953 in the training cohort, 0.930 in the internal validation cohort, 0.924 and 0.903 in the two external validation cohorts, respectively. RDLCSM performs better than RDLM (AUC range: 0.831-0.890) and 2DLCSM (AUC range: 0.785-0.916) based on KNN. In addition, RDLCSM had the highest accuracy (0.818-0.882) and specificity (0.926-1.000).

INTERPRETATION

The RDLCSM, which combines whole-tumor radiomics, 2D and 3D deep learning, clinical-visual radiology, and subregional omics, can be used as a non-invasive tool to predict thymoma risk classification.