Prediction of Postoperative Recurrence of Vocal Cord Leukoplakia Based on Multiscale Features of Laryngoscopic Images.

BACKGROUND AND OBJECTIVE: Vocal cord leukoplakia (VCL) is a common precancerous lesion of the larynx associated with a high postoperative recurrence rate. Current prognostic assessments primarily rely on subjective clinical features, and quantitative prediction models based on deep information from laryngoscopic images remain scarce. This study aims to extract quantitative image features from laryngoscopic images to develop and validate a predictive model for 2-year postoperative recurrence of VCL and to identify key prognostic features. METHODS: A total of 241 patients with VCL from two centers were retrospectively enrolled. Regions of interest covering the leukoplakia lesions were delineated on laryngoscopic images and transformed into multiple color spaces. A comprehensive set of quantitative image features was extracted, including texture, frequency domain, and shape characteristics. Feature selection was performed using statistical correlation analysis and machine learning‑based algorithms. Ten machine learning algorithms were employed for model construction. Model performance was evaluated using metrics, including the area under the receiver operating characteristic curve (AUC). An interpretability method (SHapley Additive exPlanations [SHAP]) was used to explain which features contributed most to the model's predictions. RESULTS: Ten core features were ultimately retained after selection. The support vector machine model achieved the best performance on the test set, with an AUC of 0.87. Calibration curves and decision curve analysis confirmed its satisfactory predictive reliability and clinical net benefit. SHAP analysis ranked the importance of the selected features. CONCLUSION: This study suggests that quantitative image features derived from laryngoscopic images may serve as potential indicators for predicting postoperative recurrence of VCL. The proposed model, together with the identified features, may offer complementary insight for prognostic assessment and warrant further validation in future studies.