Laryngoscope8: Laryngeal image dataset and classification of laryngeal disease based on attention mechanism

Laryngeal disease is a common disease worldwide. However, currently there are no public laryngeal image datasets, which hinders the development of automatic classification of laryngeal disease. In this work, we build a new laryngeal image dataset called Laryngoscope8, which comprises 3057 images of 1950 unique individuals, and the images have been labeled with one of eight labels (including seven pathological labels and one normal label) by professional otolaryngologists. We also propose a laryngeal disease classification method, which uses attention mechanism to obtain the critical area under the supervision of image labels for laryngeal disease classification. That is, we first train a CNN model to classify the laryngeal images. If the classification result is correct, the region with strong response is most likely a critical area. The regions with strong responses are used as training data to train an object localization model that can automatically locate the critical area. Given an image for classification, the trained object localization model is employed to locate the critical area. Then, the located critical area is employed for image classification. The entire process only requires image-level labels and does not require manual labeling of the critical area. Experiment results show that the proposed method achieves promising performance in laryngeal disease classification.