研究目的：青少年特发性脊柱侧凸(Adolescent Idiopathic Scoliosis, AIS)患者术后外观改善与生活质量提升密切相关，传统随访主要依赖 X 线检查，存在辐射风险和资源受限等问题。为此，基于临床随访数据，探讨美学参数变化与 SRS-22r 评分的相关性，构建深度学习模型，实现术前术后关键点识别与多模态结构预测，我们提出以 2D-RGB 图像生成虚拟 X 线影像的方法，旨在建立一个融合主观感受、外观参数与智能预测的AIS术后评估的尝试。

研究方法：本研究由三部分组成。第一部分为临床数据回顾分析。纳入2019年2月至2021年11月于北京协和医院接受后路脊柱融合术(Posterior Spinal Fusion, PSF)并完成2年随访的79例重度AIS患者，收集术前及术后背部标准化 2D-RGB 图像、SRS-22r 评分及相关临床资料。使用 Image-Pro 软件测量肩部、颈部、躯干和腰线等部位的美学参数，并以 SRS-22r 各领域最小临床重要差异(Minimal Clinically Important Difference, MCID)为标准对患者分组分析美学变化与生活质量改善之间的关联。第二部分基于关键点识别美学分析模型的构建。回顾性收集2018年1月至2024年10月间来自四家大型医院共685例 AIS 患者的 PSF 术前术后 2D-RGB 图像与全脊柱正位 X 线资料。采用高分辨率网络提取图像关键点，结合多模态整合感知机构建两阶段模型，识别并预测患者术前术后美学角度、Cobb 角与主弯类型，实现美学与结构信息的联合建模；通过 PCK、R²、MAE、MSE、F1-score 等指标在训练集、内部测试集和外部测试集进行模型性能的系统评估，验证其准确性与泛化能力。第三部分构建条件生成对抗网络 XR-GAN，融合 CycleGAN 与 pix2pix 结构，输入术后 2D-RGB 背部图像，输出匹配的虚拟术后全脊柱正位 X 线图像。模型集成自注意机制以提升空间结构还原能力，采用 U-Net 生成器与 CNN 判别器增强图像质量。训练集包含来自 PUMCH 的992对图像，内部测试集110对，外部测试集127对，合计1229对配对图像。模型评估指标包括术后 Cobb 角预测的 MAE、RMSE、R² 及图像质量评价指标(FID、SSIM、PSNR、LPIPS)与专家打分 (ER)，比较其在结构准确性和视觉真实感上的表现。

研究结果：第一部分，79例接受 PSF 的 AIS 患者在术后美学参数中，MS、LS、NA、ATR、SA、WD 及 WH 均有显著改善(p<0.05)。术后 SRS-22r 评分达到 MCID 的患者在特定美学参数(NA、WH、WD)上的改善幅度优于未改善者。Logistic 回归结果表明，MS、ATR 和 WD 的改善分别是总分、功能和疼痛评分改善的主要预测因子。第二部分，在内部测试集中，术前关键点识别准确，3像素阈值下PCK为83%（LSL）至94%（AAL），术后略低，WL、SAL 等区域表现不佳。术前美学角度回归表现优良，R² 均值为0.931（最高为 MS：R²=0.966，MAE=0.33°），术后阶段略有下降，MS 受影响显著（R² 降至0.771）。外部测试集整体精度较低，角度 R² 均值为0.894，MAE 上升至0.51°。在主弯类型分类中，T 型识别准确率高，TL 型在外部测试集中 precision 仅0.167。结果表明，模型在内部数据上表现稳定，但对术后外观变化与异构数据的适应性仍有限，泛化能力有待进一步提升。第三部分在内部测试集中，XR-GAN 的 Cobb 角预测MAE为1.54，RMSE 为1.97，R² 达0.964，Pearson 相关系数为0.983，远超 Pix2Pix、CycleGAN 等对比模型。在外部测试集中，XR-GAN 保持较高性能(R² = 0.873，Pearson R = 0.937)。分级预测方面，XR-GAN在术后Cobb角严重程度识别上准确率达0.836（内部）与0.717（外部），对 Grade IV(>20°)几乎无误判。在图像质量评价中，XR-GAN 在 SSIM、PSNR、LPIPS 等指标上均优于其他模型（如 SSIM 达0.920，LPIPS 仅0.188）。

结论：本研究围绕 AIS 术后评估构建了由现象观察、智能预测到远程随访新选择的递进路径。首先基于真实临床随访数据，明确了美学参数改善（如肩部平衡、腰线对称、躯干旋转）与生活质量提升之间的显著关联，突出了背部美学在术后评估中的临床价值；其次，构建 HRNet 模型，识别及预测术前术后关键点并量化美学参数，尝试个体外观特征与影像学指标的同步预测；进一步提出 XR-GAN 模型，通过 2D-RGB 图像生成虚拟全脊柱正位 X 线，实现术后结构信息的无创、高保真重建，具备临床应用潜力。

Objective: Postoperative appearance improvement in adolescents with idiopathic scoliosis (AIS) is closely linked to better quality of life. Traditional follow-up relies heavily on X-ray imaging, which poses radiation risks and resource limitations. This study analyzes the correlation between aesthetic changes and SRS-22r scores, develops deep learning models for keypoint detection and multimodal structural prediction, and proposes a 2D RGB-based method to generate synthetic full-spine X-ray images for safer follow-up.

Methods: This study comprises three components. First, a retrospective analysis was conducted on 79 severe AIS patients who underwent PSF at PUMCH (2019–2021), with standardized pre- and postoperative 2D-RGB images, SRS-22r scores, and clinical data collected. Cosmetic parameters were measured, and patients were grouped based on the MCID of SRS-22r to analyze the association between cosmetic improvements and quality of life. Second, a two-stage deep learning model was developed using data from 685 AIS patients across four hospitals (2018–2024). HRNet was used for keypoint detection, and a multimodal integration MLP (MMI-MLP) was designed to predict pre- and postoperative cosmetic angles, Cobb angles, and curve types. The model’s accuracy and generalization were evaluated using PCK, R², MAE, MSE, and F1-score across training, internal, and external datasets. Third, a conditional generative adversarial network (XR-GAN), combining CycleGAN and pix2pix architectures, was built to synthesize full-spine postoperative X-rays from RGB images. The model used a U-Net generator, CNN discriminator, and self-attention mechanisms. It was trained and tested on 1,229 image pairs and evaluated using Cobb angle prediction metrics (MAE, RMSE, R²) and image quality indices (FID, SSIM, PSNR, LPIPS, ER).

Results: In the first part, 79 AIS patients who underwent PSF showed significant postoperative improvements in cosmetic parameters including MS, LS, NA, ATR, SA, WD, and WH (p < 0.05). Patients who reached the MCID threshold in SRS-22r scores exhibited greater improvements in NA, WH, and WD. Logistic regression identified MS, ATR, and WD improvements as significant predictors of total, function, and pain score improvements, respectively. In the second part, internal testing showed high accuracy in preoperative keypoint detection, with PCK values ranging from 83% (LSL) to 94% (AAL) at the 3-pixel threshold. Postoperative keypoint prediction was less accurate, particularly for WL and SAL. Regression performance for preoperative cosmetic angles was strong (mean R² = 0.931; best for MS: R² = 0.966, MAE = 0.33°), while postoperative results slightly declined, with MS dropping to R² = 0.771. External test performance was lower overall (mean R² = 0.894; MAE increased to 0.51°). In curve type classification, T-type was well recognized, but TL-type precision dropped to 0.167 in the external test set. In the third part, XR-GAN achieved superior Cobb angle prediction in the internal test set (MAE = 1.54, RMSE = 1.97, R² = 0.964, Pearson R = 0.983), outperforming Pix2Pix and CycleGAN. In external testing, it maintained good performance (R² = 0.873, R = 0.937). Grade classification accuracy reached 0.836 (internal) and 0.717 (external), with nearly no misclassification for Grade IV (>20°). Image quality metrics such as SSIM (0.920) and LPIPS (0.188) also outperformed other models.

Conclusion: This study established a progressive framework for postoperative evaluation of AIS, spanning clinical observation, intelligent prediction, and remote follow-up. First, real-world clinical data confirmed a significant association between improvements in cosmetic parameters (e.g., shoulder balance, waistline symmetry, trunk rotation) and enhanced quality of life, highlighting the clinical relevance of back aesthetics. Second, the HRNet-based model enabled accurate identification and prediction of pre- and postoperative keypoints, allowing quantitative assessment of cosmetic features and joint inference with structural indicators. Finally, the proposed XR-GAN model generated high-fidelity, noninvasive full-spine X-ray images from 2D-RGB inputs, offering a novel solution for postoperative structural reconstruction with promising clinical applicability.