Deep learning-based prediction of thyroid cartilage invasion: Analysis on CT images in laryngeal and hypopharyngeal squamous cell carcinoma

To provide new insights for the development of a deep learning-optimized radiotherapy assistance system, we proposed a deep learning-based model for evaluating the thyroid cartilage invasion, and explored the performance of this model by analyzing non-contrast-enhanced computed tomography (CT) images. This retrospective study included a total of 286 patients with laryngeal and hypopharyngeal squamous cell carcinoma (LHSCC) who underwent preoperative non-contrast CT scans from January 2012 to November 2022 for model training (Dataset A), as well as from November 2022 to May 2023 for validation (Dataset B). 3D CT images were cropped to cover the entire cartilage. The ResNet-3Dsml model, adapted from the ResNet architecture for binary classification, was used for prediction of thyroid cartilage invasion. The deep learning model shows predictive performance (AUC 0.844/0.856) similar to that of radiologists on Dataset A and B, demonstrating predictive accuracy of 88.3%/80.0%, specificity of 96.7%/84.0%, and sensitivity of 56.2%/40.0%. Our model exhibited high specificity but low sensitivity, resembling the diagnostic pattern of radiologists. The ResNet-3Dsml model, based on non-contrast 3D CT images, showed high AUC in predicting thyroid cartilage invasion in patients with LHSCC, offering a cost-effective and minimally invasive new assessment method for clinical practice.