Development and validation of a deep learning radiomics nomogram for preoperatively differentiating thymic epithelial...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Thymic epithelial tumors (TETs) require accurate preoperative risk assessment for optimal treatment planning.
• Current methods may not fully capture the complexity of TETs for risk stratification.

Purpose of the Study:

• To develop and validate a deep learning radiomics nomogram (DLRN) for predicting the preoperative risk status of patients with thymic epithelial tumors (TETs).

Main Methods:

• A transformer-based convolutional neural network was used to extract deep learning features from CECT scans of 257 TET patients.
• A deep learning signature (DLS) was created, and a DLRN was developed incorporating DLS, clinical characteristics, and subjective CT findings.
• Model performance was evaluated using receiver operating characteristic curves and AUC values across training and validation cohorts.

Main Results:

• The DLRN, combining subjective CT features and DLS, demonstrated high performance in differentiating TET risk status.
• Area Under the Curve (AUC) values reached 0.959 in the training cohort and ranged from 0.846 to 0.868 in validation cohorts.
• The DLRN proved to be the most predictive and clinically useful model compared to DLS, radiomics signature, or clinical models.

Conclusions:

• A DLRN integrating CECT-derived deep learning features and subjective findings offers a high-performance, non-invasive method for predicting TET risk status.
• This approach can aid in preoperative risk stratification, prognostic evaluation, and personalized therapy decisions for TET patients.