A Dual-Branch Contrastive Network Using Symmetric Counterfactual Generation for Multi-Modal Ocular Adnexal Lymphoma Segmentation

Precise segmentation of ocular adnexal lymphoma aids physicians in subsequent tumor classification and diagnosis. However, ocular adnexal lymphoma presents significant challenges for tumor segmentation due to its variable location and irregular shape. Recently, many methods based on counterfactual image generation have been proposed to restore lesion images to pseudo-healthy images to assist medical image segmentation. However, since paired lesion-healthy images are difficult to obtain, most methods rely on unsupervised, unpaired approaches to generate healthy images. Since the human head structure exhibits symmetry, and the ocular adnexal lymphoma we are studying primarily affects a single orbital region located on the head, we utilize the symmetry of the human eye structure to divide it into two parts, thereby obtaining paired lesion-healthy images for image generation. We propose a Mamba-based dual-branch network architecture for multi-modal ocular adnexal lymphoma segmentation, using the generated pseudo-healthy images as a reference to assist tumor segmentation. This approach effectively improves segmentation performance and outperforms state-of-the-art segmentation methods.