Unsupervised Domain Adaptation Approach for Liver Tumor Detection in Multi-phase CT Images

Abstract

AbstractFor computer-aided diagnosis, automatic and accurate liver tumor detection in multi-phase CT images is essential. Nowadays, deep learning has been widely used in various medical applications. Deep learning-based AI systems require a large amount of training data for model learning. However, acquiring sufficient training data with high-quality annotation is a major challenge in Healthcare. As a result, deep learning-based models face a lack of annotated training data problem. While the generalization of a label-rich training domain (source) to a new test domain (target) causes a domain shift problem in deep learning-based models. To solve the lack of training data and domain shift problem, domain adaptation-based methods have recently been developed as a technique to bridge the domain gap across datasets with different feature characteristics and data distributions. In this chapter, we have proposed domain adaptation-based technique for liver tumor detection in multi-phase CT images. We discuss the domain-shift problem in different phases of multiphase liver CT images and introduce our domain adaptation technique for multi-phase CT images. We have used PV phase images to learn a model and applied the learnt model to ART and NC phase images (i.e. different domains) by adapting the domain knowledge. To address the domain gap between the different phases of CT images, we employ adversarial learning scheme using an anchor-free object detector. Further, we propose to use the maximum square loss for mid-level output feature maps to improve the performance. Our method does not require separate object-level annotations for each phase of Multiphase CT image while training. The results of the experiments show that models trained using our proposed domain adaptation technique perform much better than those trained in normal setting.KeywordsLiver tumor detectionMulti-phase ct imageAnchor-free object detectorsDomain adaptationAdversarial learningMaximum square loss