DCLNet: Data Closed-Loop Network for Laryngoscopy Image Annotation and Classification

Laryngeal diseases diagnosis is a common challenge in medical image processing. Traditional methods usually use classification networks to process entire images or their cropped key area images. However, the former may not effectively focus on the crucial region within the image, while the latter is a two-stage task that cannot utilize global information. To address this problem, we propose a novel Data Closed-Loop Network (DCLNet) for laryngoscope image classification, fully utilizing the global features in the image and enabling end-to-end training. By introducing attention mechanisms, DCLNet can automatically expand the Laryngoscope 8 classification dataset into a high-quality annotated object detection dataset. Moreover, to address some inherent issues of medical datasets, such as poor interpretability and a long-tailed distribution, we employ the methods of data closed-loop and curriculum learning. In this way, performance on the baseline dataset can be improved by optimizing the annotation of the training set without changing the model structure. Extensive experiments prove that our model achieves better performance compared to existing traditional methods.