Anatomical-Persistent Synergy Mamba for Joint Pre-to-Postoperative CT Synthesis and Multi-structure Segmentation in Orbital Decompression Surgery.

Abstract

Thyroid Eye Disease (TED) often causes proptosis and optic nerve compression, and orbital decompression surgery is commonly performed to relieve these symptoms. However, postoperative anatomical changes are highly nonlinear and patient-specific due to bone removal and orbital content displacement, which makes reliable preoperative prediction both critical and challenging. Existing methods for pre- to postoperative medical image generation often overlook an important clinical prior that, although surgery induces local structural changes, the overall anatomy of the same patient remains largely consistent, and neglecting this can lead to anatomically implausible deviations. To address this issue, we propose APS-Mamba (Anatomical-Persistent Synergy Mamba), a conditional multi-task framework that takes preoperative CT and surgical information as input and jointly performs postoperative CT synthesis and multi-structure segmentation. The model introduces an anatomical persistence constraint to preserve structural consistency in non-operated regions while allowing realistic local deformations in surgery-related areas. It also incorporates a generation-segmentation synergistic learning strategy. In this design, morphological image synthesis provides predicted variations for segmentation, and segmentation enforces semantic constraints to improve the structural fidelity of generated images. Experiments on a clinical TED dataset demonstrate that APS-Mamba outperforms state-of-the-art methods in both image synthesis (PSNR 28.60 dB, SSIM 91.72%) and segmentation (mean DSC 86.17%). Ablation studies further confirm the effectiveness of the anatomical constraint and synergistic learning strategy in modeling anatomical changes.