Sendecki Adam, Ledwoń Daniel, Tuszy Aleksandra, Nycz Julia, Wąsowska Anna, Boguszewska-Chachulska Anna, Mitas Andrzej W, Wylęgała Edward, Teper Sławomir
Chair and Clinical Department of Ophthalmology, Faculty of Medical Sciences in Zabrze, Medical University of Silesia, 40-752 Katowice, Poland.
Faculty of Biomedical Engineering, Silesian University of Technology, 41-800 Zabrze, Poland.
Biomedicines. 2024 Sep 13;12(9):2092. doi: 10.3390/biomedicines12092092.
Age-related macular degeneration (AMD) is a complex eye disorder with an environmental and genetic origin, affecting millions worldwide. The study aims to explore the association between retinal morphology and the polygenic risk score (PRS) for AMD using fundus images and deep learning techniques.
The study used and pre-processed 23,654 fundus images from 332 subjects (235 patients with AMD and 97 controls), ultimately selecting 558 high-quality images for analysis. The fine-tuned DenseNet121 deep learning model was employed to estimate PRS from single fundus images. After training, deep features were extracted, fused, and used in machine learning regression models to estimate PRS for each subject. The Grad-CAM technique was applied to examine the relationship between areas of increased model activity and the retina's morphological features specific to AMD.
Using the hybrid approach improved the results obtained by DenseNet121 in 5-fold cross-validation. The final evaluation metrics for all predictions from the best model from each fold are MAE = 0.74, MSE = 0.85, RMSE = 0.92, R = 0.18, MAPE = 2.41. Grad-CAM heatmap evaluation showed that the model decisions rely on lesion area, focusing mostly on the presence of drusen. The proposed approach was also shown to be sensitive to artifacts present in the image.
The findings indicate an association between fundus images and AMD PRS, suggesting that deep learning models may effectively estimate genetic risk for AMD from retinal images, potentially aiding in early detection and personalized treatment strategies.
年龄相关性黄斑变性（AMD）是一种具有环境和遗传起源的复杂眼部疾病，影响着全球数百万人。本研究旨在利用眼底图像和深度学习技术探索视网膜形态与AMD多基因风险评分（PRS）之间的关联。
本研究使用并预处理了来自332名受试者（235例AMD患者和97名对照）的23654张眼底图像，最终选择558张高质量图像进行分析。使用微调后的DenseNet121深度学习模型从单张眼底图像估计PRS。训练后，提取、融合深度特征，并将其用于机器学习回归模型以估计每个受试者的PRS。应用Grad-CAM技术检查模型活动增加区域与AMD特有的视网膜形态特征之间的关系。
在五折交叉验证中，使用混合方法改善了DenseNet121获得的结果。来自每一折最佳模型的所有预测的最终评估指标为：平均绝对误差（MAE）=0.74，均方误差（MSE）=0.85，均方根误差（RMSE）=0.92，相关系数（R）=0.18，平均绝对百分比误差（MAPE）=2.41。Grad-CAM热图评估表明，模型决策依赖于病变区域，主要关注玻璃膜疣的存在。所提出的方法也被证明对图像中存在的伪影敏感。
研究结果表明眼底图像与AMD PRS之间存在关联，这表明深度学习模型可能有效地从视网膜图像估计AMD的遗传风险，可能有助于早期检测和个性化治疗策略。
