Semi-supervised deep learning for automatic classification of medical images

In recent years, the prevalence rate of laryngeal cancer is gradually rising, and the best treatment at present lies in early diagnosis and prevention. Early laryngeal cancer is difficult to be detected, and imaging examination is one of the important methods for diagnosis and prevention of laryngeal cancer. However, due to the particularity of medical images, the human eye is often difficult to pay attention to all characteristic information, and doctors lacking experience in endoscopy are prone to misjudgment. Using deep learning to build a computer-aided diagnosis system for laryngeal diseases is an effective solution. At present, semi-supervised learning has become a research hotspot in the field of medical image classification. The intent of this study is to construct and evaluate the performance of a technique based on self-training and consistent regularization for the classification of laryngeal disease images. Our dataset contains a total of 4000 laryngeal images, including 3000 normal images and 1000 laryngeal cancer images. We divide the sample set into two parts: learning samples and validation samples. The distribution ratio is 8:2. Thirty percent of the samples were labeled and seventy percent were unlabeled. Experimental data show that the classification accuracy of the model can reach 80.125%. The classifier is helpful to develop a computer-aided diagnosis system for laryngeal diseases and can effectively decrease the annotation cost.