Projection-based tokenization with Pseudo feature learning for ovarian lesion segementation in ultrasound images

Accurate segmentation of ovarian lesions in ultrasound images is essential for early cancer detection and effective treatment planning. However, this task remains challenging due to large variations in lesion size and shape, as well as similar intensities or textures between lesions and surrounding tissues. To jointly address these challenges, this study proposes a Radon-projected model with pseudo feature learning for ovarian lesion segmentation. Specifically, a Radon-projected feature learning module generates projection-based tokenization as input to the Transformer and integrates both local and global information, thereby preserving global structural correlations and improving adaptability to lesions of varying sizes and shapes. Furthermore, a pseudo feature learning module constructs pseudo-foreground and pseudo-background regions, and semantic correlations within and between these regions are explored via attention mechanisms, thus enhancing feature discrimination and mitigating challenges arising from similar intensities and textures. Extensive experiments on three ovarian and two non-ovarian ultrasound datasets demonstrate that our approach achieves superior segmentation performance and strong generalizability.