Missed diagnoses detection by adversarial learning

Abstract

Missed diagnosis has been a serious public health issue in clinical diagnosis and treatment, it may cause disease deterioration and reduce cure rate. In recent years, many deep learning approaches have been proposed for automated medical image classification. In this study, we propose a new methodology to detect the images missed diagnosed by deep learning classifiers. Based on the intermediate feature maps of deep learning classifiers, the proposed model can detect missed diagnosed sample and reduce the missed diagnoses rate, with no obvious decrease in the accuracy. The proposed model is constructed using generative adversarial networks and autoencoders to learn consistent mapping from data space to latent space, and is trained with adversarial examples. After training, the output of the discriminator is used to recognize missed diagnosed samples. The method is evaluated on different network architectures and various types of medical image datasets and achieves promising results. Compared with other state-of-the-art approaches, the proposed method shows superior performance on most datasets.