Abstract: Abstract: An innovative "segmentation-prediction" cascade framework was proposed for non-invasive prediction of coronary heart disease and multi-vascular disease risks, leveraging retinal vessel imaging. The MA_DNet framework, incorporating ResMamba and GMSF modules, significantly improved vascular segmentation accuracy and continuity. The model utilizes a two-stage approach: In the first stage, vessel features (density and fractal dimension) and clinical data were combined to predict coronary heart disease risk and generate intermediate variables for multi-vascular disease prediction. In the second stage, enhanced features were used to achieve precise predictions for multi-vascular diseases. Experimental results showed that the MA_DNet framework achieved Dice coefficients of 93.02% and Jaccard indices of 87.04% on a subset of the OCTA-500 dataset, improving 0.28% and 0.37% over the IPN_V2+ model, respectively. On another subset of the same dataset, the Dice coefficient and Jaccard index reached 89.76% and 81.52%, improving 0.35% and 0.57%, respectively. Moreover, the disease prediction model’s AUC improved from 0.57 to 0.79. This study is the first to combine retinal vessel segmentation with systemic disease prediction, offering a novel, interpretable framework for non-invasive cardiovascular and cerebrovascular disease diagnosis and validating the efficacy of cross-modal feature correlation.

Key words: Keywords: coronary heart disease prediction, multi-vascular disease, retinal vessel segmentation, deep learning, medical image analysis, non-invasive diagnosis, optical coherence tomography angiography (OCTA)

摘要： 摘 要: 提出一种视网膜血管影像解析与疾病预测协同框架。基于三维分割网络MA_DNet（融合ResMamba与GMSF模块），提升血管结构分割精度及连续性。创新性建立"分割-预测"级联范式：首阶段结合血管特征（密度、分形维数）与临床数据预测冠心病风险并生成中间变量，次阶段利用增强特征实现多血管疾病精准预测。构建的MA_DNet框架在OCTA-500的一个数据子集上的Dice系数和Jaccard指数分别达到93.02%和87.04%，较IPN_V2+提升0.28%和0.37%；在OCTA-500的另一子集上分别为89.76%和81.52%，提升0.35%和0.57%；疾病预测模块AUC从0.57提升至0.79。该研究首次实现视网膜血管分割与系统性疾病的级联预测，通过可解释性计算框架为心脑血管疾病无创诊断提供新思路，验证了跨模态特征关联的有效性。

关键词: 关键词: 冠心病预测, 多血管疾病, 视网膜血管分割, 深度学习, 医学影像分析, 非侵入性诊断, 光学相干断层扫描