Unsupervised cross-domain semantic segmentation on multi-modality ovarian tumor ultrasound data

Ovarian cancer is one of the most harmful gynecological diseases. Early detection of ovarian tumors, facilitated by computer-aided techniques, is key to significantly reducing mortality rates. Amidst advancements in medical diagnostics, ultrasound imaging has emerged as a ubiquitous tool in clinical environments. Nevertheless, current methods mainly focus on single-modality ultrasound segmentation or recognition of ovarian tumors, overlooking the representation potential of multi-modality imaging. To address this problem, we propose a Multi-Modality Ovarian Tumor Ultrasound (MMOTU) image dataset containing 1469 2d-ultrasound images and 170 contrast-enhanced ultrasonography (CEUS) images with pixel-wise and global-wise annotations. We focus on an unsupervised cross-domain semantic segmentation task to explore the model's adaptation potential from 2d-ultrasound and CEUS modalities. For this task, we propose a feature alignment-based architecture named Dual-Scheme Domain-Selected Network (DS²Net). Specifically, we embed adversarial learning to conduct feature-level alignment and propose the Domain-Distinct Selected Module (DDSM) and Domain-Universal Selected Module (DUSM) to represent the distinct and universal features in source-style and target-style. Extensive experiments and analysis on the MMOTU image dataset underscore the remarkable performance of DS²Net in fostering bidirectional cross-domain adaptation between 2d-ultrasound and CEUS images, thereby enhancing segmentation accuracy and pushing the boundaries of ovarian tumor detection technologies. Our proposed dataset and code are available at https://github.com/cv516Buaa/MMOTU_DS2Net.