Radiomics-based unsupervised clustering and deep learning nomogram for response prediction after neoadjuvant chemotherapy in locally advanced laryngeal cancer

OBJECTIVES: To develop a nomogram model for individualized prediction of neoadjuvant chemotherapy (NAC) response in locally advanced laryngeal cancer (LALC). METHODS: A total of 175 patients who underwent CT examinations before NAC from two hospitals were retrospectively enrolled and divided into training (n = 112) and test (n = 63) sets. Significant radiomics and clinical features were selected sequentially, and consensus clustering was used to identify tumor subtypes. A nomogram model for predicting remission status was constructed by fusing clinical signature, radiomics signature, cluster result and deep learning signature. The predictive performance of junior and senior doctors for NAC response was evaluated with and without model assistance. RESULTS: Two radiomics subtypes (Subtype A and B) were identified. Compared with Subtype A, Subtype B had higher frequencies of heterogeneous tumor density (all P < 0.001) and marked venous-phase enhancement (all P < 0.05) in both sets. The nomogram-derived remission score showed good predictive performance for NAC response, with AUCs of 0.935 (training set) and 0.912 (test set). Survival analysis revealed that patients with high remission scores had better overall survival (OS) and locoregional control than those with low scores (all P < 0.05) in both sets. Model assistance improved the predictive performance of junior (AUC: 0.799 vs. 0.915) and senior doctors (AUC: 0.849 vs. 0.919, all P < 0.05). CONCLUSION: The nomogram model based on two-center databases achieved good performance in predicting NAC response and survival in LALC patients, which may contribute to the personalized treatment of LALC.