Deep learning-based radiomic nomogram to predict risk categorization of thymic epithelial tumors: A multicenter study

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Thymic epithelial tumors (TETs) require accurate risk stratification for effective treatment planning and prognosis.
• Current methods for risk categorization can be improved with advanced predictive tools.

Purpose of the Study:

• To develop and evaluate a deep learning-based radiomics model for predicting the histological risk categorization of TETs.
• To create a visual tool, a deep learning-based radiomic nomogram (DLRN), for assessing classification performance.

Main Methods:

• Utilized preoperative contrast-enhanced CT images from 681 patients across three hospitals.
• Extracted handcrafted and deep learning features to develop radiomics signatures (Rad_Sig, DL_Sig, DLR_Sig).
• Compared model performance using receiver operating characteristic and decision curve analysis (DCA).

Main Results:

• The deep learning radiomics signature (DLR_Sig) showed promising performance (AUC 0.883 derivation, 0.749 external test).
• The DLRN, combining DLR_Sig with age and gender, achieved superior performance (AUC 0.965 derivation, 0.786 external test).
• DLRN demonstrated greater clinical utility compared to other radiomics signatures in DCA.

Conclusions:

• A DLRN was successfully developed and validated for accurate TET risk prediction.
• The DLRN shows potential for facilitating individualized treatment strategies.
• This tool can enhance the evaluation of patient prognosis in TET cases.