Enhanced Liver Lesion Detection in Multi-Phase CT Images Using Unsupervised Domain Adaptation and Feature Generalization

For computer-aided diagnosis, automatic and precise detection of liver lesions in multi-phase CT images is important. Recently, deep learning has emerged as a prominent tool in various computer vision tasks including medical applications, offering promising solutions. However, the utility and scalability of deep learning-based AI systems heavily rely on large quantities of well-annotated training data. The lack of annotated training data poses significant challenges for developing robust models. Additionally, the distribution difference or data drift from a training domain (source) to an unseen test domain (target) presents a domain shift problem in deep learning models. In medical imaging, the inherent differences among multiphase images and variations in parameters across different imaging centers during image acquisition create major challenges for predictive models. To address the lack of training data and the domain shift issue, domain adaptation-based methods have emerged as a viable solution to reduce the domain gap across datasets with distinct feature characteristics and data distributions. In this chapter, we discuss domain adaptation-based techniques for liver tumor detection in multi-phase CT images. To bridge the domain gap between multiple centers and different phases of CT images, we present an adversarial learning and feature generalization scheme using an anchorfree object detector. We propose to integrate divergence based feature adaptive techniques to generalize the framework. Our method significantly increases the detection accuracy and achieves state-of-the-art performance using CT images from different centers and different phases.