CT-based radiomics and deep learning models for predicting thyroid cartilage invasion and patient prognosis in...

Accurate assessment of thyroid cartilage invasion is crucial for treatment decision-making and prognosis evaluation in laryngeal squamous cell carcinoma (LSCC). This study aimed to compare the performance of the radiomics and deep learning (DL) models for predicting thyroid cartilage invasion in LSCC patients, and evaluate prognostic value of the optimal predictive model. A total of 418 pathologically confirmed LSCC patients from two centers were enrolled and divided into a training cohort (n = 247), an internal validation cohort (n = 110), and an external validation cohort (n = 61). Models were developed based on venous-phase CT images and compared with two radiologists. A nomogram incorporating the optimal model and clinical risk factors was also constructed. Additionally, the prognostic value of the optimal model was assessed regarding disease-free survival (DFS). The 2D DL model showed better performance in predicting thyroid cartilage invasion, and the corresponding nomogram integrating 2D DL signature and clinical risk factors achieved the highest AUCs. However, no differences in AUCs were found in the external validation cohort (p > 0.05 for all). Additionally, the 2D DL signature and clinical N stage were independent predictors of DFS. The 2D DL-based nomogram demonstrated satisfactory predictive performance for thyroid cartilage invasion and prognosis in patients with LSCC.