ClusterUDA: Latent Space Clustering in Unsupervised Domain Adaption for Pulmonary Nodule Detection

Deep learning has achieved notable performance in pulmonary nodule (PN) detection. However, existing detection methods typically assume that training and testing CT images are drawn from a similar distribution, which may not always hold in clinic due to the variety of device vendors and patient population. Hence, the idea of domain adaption is introduced to address this domain shift problem. Although various approaches have been proposed to tackle this issue, the characteristics of samples are ignored in specific usage scenarios, especially in clinic. To this end, a novel unsupervised domain adaption method (namely ClusterUDA) for PN detection is proposed by considering characteristics of medical images. Specifically, a convenient and effective extraction strategy is firstly introduced to obtain the Histogram of Oriented Gradient (HOG) features. Then, we estimate the similarity between source domain and target one by clustering latent space. Finally, an adaptive PN detection network can be learned by utilizing distribution similarity information. Extensive experiments show that, by introducing a domain adaption method, our proposed ClusterUDA detection model achieves impressive cross-domain performance in terms of quantitative detection evaluation on multiple datasets. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.