Deep-learning Classifier Differentiates Voice States from Stroboscopy

A deep-learning classifier successfully differentiated between healthy vocal folds (HVF), unilateral vocal fold paralysis (UVFP), and vocal fold lesions (benign/malignant) using voice data from videolaryngostroboscopy. The binary classifier achieved an accuracy of 83% and F1-score of 0.90 on the test set. However, multi-class classification performance was weaker, with 40% accuracy and 0.36 F1-score. These findings suggest deep learning can aid in diagnosing vocal fold conditions, with better performance in binary classifications.