Segmentation-Driven Deep Learning for Explainable and Automated Vocal Fold Disorder Classification

The abnormalities of the vocal folds cause voice disorders that affect speech and communication. The traditional diagnostic tools such as laryngoscopy and high-speed video endoscopy are subjective and time consuming and may be affected by inter-observer variation. In order to overcome these difficulties, we suggest a deep-learning-based paradigm of automatic vocal-fold segmentation and diagnosis of disorder on the Emory Vocal Folds dataset (approximately 2.9k of annotated images). We use preprocessing (cleaning, normalization and augmentation, class balancing), semantic segmentation as U-Net, and feature extractions as state-of-the-art CNNs (ResNet, DenseNet, VGG16, MobileNet and EfficientNet). The extracted features were used with both the end-to-end CNN classification and traditional machine learning classifiers (SVM, Random Forest, Gradient Boosting). It was experimentally demonstrated that EfficientNet was the most accurate at validation (98%), as compared to ResNet, DenseNet, and VGG16. Classical models also showed good results with an accuracy of 97% with random forest and 96 with SVM. The interpretability of the models was guaranteed through Grad-CAM, Grad-CAM++, Integrated Gradients, and Occlusion Sensitivity, and it was confirmed that the predictions paid attention to the parts of the vocal folds that are clinically relevant. These results indicate that deep learning can offer a robust, interpretable, and non-invasive algorithm for the early detection and diagnosis of vocal fold disorders thereby providing a possibility of real-time clinical decision support and mobile health diagnosis.