A Multimodal Approach for Deep‐Learning Classification of Vocal Fold Pathologies in Stroboscopy

OBJECTIVE: To develop and validate a multimodal deep-learning classifier trained on stroboscopic image, voice, and clinicodemographic data, differentiating between three different vocal fold (VF) states: healthy (HVF), unilateral paralysis (UVFP), and VF lesions, including benign and malignant pathologies. METHODS: Patients with UVFP (n = 54), VF lesions (n = 42), and HVF (n = 41) were retrospectively identified. Image frames and voice samples were extracted from stroboscopic videos. Clinicodemographic variables were collected from the electronic health record. Patient-level data was independently divided into training (80%) and testing (20%). Visual features were extracted using a transformer DINOv2 and acoustic features were extracted using Librosa. All three feature modalities were combined using a custom multilayer perceptron. Unimodality models using only image or only voice data were trained for comparison. Accuracy and F1 scores were used to validate the models. RESULTS: On a hold-out test set, the multimodal classifier demonstrated stronger performance (76.9% accuracy) compared to the image classifier (61.5% accuracy) and audio classifier (65.4% accuracy). On an external dataset, the multimodal classifier accuracy dropped to 45%, though still an improvement compared to accuracies of 42% and 31% for the video-only and audio-only modalities, respectively. CONCLUSIONS: In this proof-of-concept study, we successfully developed a multimodal dataset and classifier for VF pathology, demonstrating the potential of combining stroboscopic frames, voice and text data. The multimodal classifier achieved higher accuracy than the image-only model and audio-only models. Future models should validate these findings on larger datasets.