Multiview Laryngoscopic Image Fusion for Patient-Level Stratification of Vocal Fold Lesions

Malignant or high-grade vocal fold disease is usually judged across repeated laryngoscopic views at different observation distances and fields of view, whereas most artificial intelligence systems remain frame-based. We evaluated whether patient-level fusion of multiview laryngoscopic image sets could improve benign versus malignant/high-risk stratification of vocal fold lesions. In this retrospective multicenter study of 628 patients, internal data from the Cancer Hospital and Institute of the Chinese Academy of Medical Sciences collected between 2008 and 2025 were split into training (n = 380), validation (n = 109), and internal test (n = 54) cohorts, and an external clinical cohort from three tertiary hospitals included 85 patients. Patient-level fusion performance was calculated in 82 external patients with complete evaluable model outputs. Lesion-centered segmentation, candidate convolutional backbones, and patient-level multiview fusion were evaluated against ensemble and multi-instance learning comparators. ResNet101 was selected as the frame-level backbone, and the 75th percentile fusion configuration showed the best generalization. In the external evaluable cohort, the final model achieved an AUC of 0.890 (95% CI, 0.823-0.957), accuracy of 0.805 (95% CI, 0.706-0.876), sensitivity of 0.643 (95% CI, 0.492-0.770), specificity of 0.975 (95% CI, 0.871-0.996), PPV of 0.964 (95% CI, 0.823-0.994), and NPV of 0.722 (95% CI, 0.591-0.824). Patient-level fusion provided the highest AUC point estimate among tested strategies, but the main clinical signal was very high specificity rather than rule-out sensitivity. At the present operating point, the model should not be used as a standalone screening or rule-out tool for deciding against biopsy; rather, it should be interpreted as an adjunct for endoscopic triage and malignant-risk stratification before biopsy.