EVENT DOSSIER
The researchers proposed an interpretable OCTA process aimed at establishing a transparent non-diagnostic connection between retinal vascular measurements and evidence in Alzheimer’s disease research. This process integrates labeled perceptual vascular segmentation, hierarchical vascular biomarker extraction, and measurement-based large language model report generation techniques. Using 117 ROSE-1 images from 39 subjects, the study segmented shallow and deep vascular complexes and their junctions, achieving a ROC-AUC recognition score ranging from 0.916 to 0.970. Six density and fractal dimension biomarkers were extracted to construct a feature spectrum, and a set of internally consistent low-density, low-fractal-dimensional phenotypes was successfully identified without labeled diagnostic labels. The generated reports were evaluated and verified using GPT, Gemini, and Llama models.
研究人员提出一种可解释的光学相干断层扫描血管造影（OCTA）流程，整合标注感知血管分割、分层血管生物标志物提取及基于测量的大型语言模型报告。该流程利用 39 名受试者的 117 张 ROSE-1 图像，对浅层血管复合体、深层血管复合体及两者结合进行分割，实现ROC-AUC值为0.916至0.970的识别效果。分析提取出六个密度和分形维数生物标志物以构建受试者水平特征谱，并在未标记诊断标签的情况下识别出一组内部一致的低密度、低分形维数表型。生成的报告经GPT、Gemini和Llama模型评估，旨在建立视网膜血管测量与阿尔茨海默病研究证据链接之间的透明非诊断连接。