doronys/synthmed-retina-hpp-sample
收藏资源简介:
SynthMed Retina — HPP I2I样本数据集包含2,500张通过图像到图像(I2I)扩散技术生成的合成视网膜眼底图像,这些图像源自人类表型项目(HPP)10K队列的真实眼底照片。每张合成图像都配有匿名化的患者元数据,包括人口统计学信息(年龄、性别)、血液检查指标(血红蛋白、胆固醇等)、血压和身体测量数据等。该数据集是为评估I2I扩散作为医学图像去识别化方法的有效性而创建的,研究发现在强度为0.5时,能显著降低再识别风险(约20%)同时保持较高的生物标志物一致性(Pearson r≈0.7)。技术实现上使用了Stable Diffusion v1-4基础模型和特定的LoRA适配器(hpp-retina),采用DDIM调度器进行50步推理,生成分辨率为512×512的图像。
The SynthMed Retina — HPP I2I Sample dataset contains 2,500 synthetic retinal fundus images generated via image-to-image (I2I) diffusion from the Human Phenotype Project (10K) cohort, paired with anonymized patient metadata. This includes demographic data (age, gender), blood test metrics (hemoglobin, cholesterol, etc.), blood pressure, and anthropometric measurements. Created to evaluate I2I diffusion as a medical image de-identification approach, the study found that at strength=0.5, it reduces top-1 re-identification rate to ~20% while maintaining high biomarker agreement (Pearson r≈0.7). Technically, it uses Stable Diffusion v1-4 with a specific LoRA adapter (hpp-retina), DDIM scheduler for 50 inference steps, producing 512×512 resolution images.




