Convolutional neural network for voice disorders classification using kymograms

The diagnosis of voice disorders typically involves examination of laryngoscopic video frames by trained experts. Videokymography (VKG) is a useful clinical tool to represent the glottal dynamics and vibratory patterns as kymographic images. In this work, a 2D Convolutional Neural Network (2D CNN) was used to classify voice disorders from kymograms. High-speed videoendoscopy (HSV) recordings of the "Benchmark for Automatic Glottis Segmentation" (BAGLS) database were used as the corpus for the voice disorders. Kymographic images were generated from this corpus. For each classification problem addressed in this work, 90% of the generated kymograms were used to train the network and the remaining 10% was used for testing its classification performance. Classification accuracies of 94.237% and 94.8% were obtained for the two cases of binary classification (healthy vs disorders, and healthy vs muscle tension dysphonia). Ternary classification (healthy vs functional vs organic disorders) of the dataset yielded an accuracy of 93.1%.