The Maersk Mc-Kinney Moeller Institute, SDU Robotics, University of Southern Denmark, Odense, Denmark.
Steno Diabetes Center Odense, Odense University Hospital, Odense, Denmark.
Transl Vis Sci Technol. 2022 Jun 1;11(6):19. doi: 10.1167/tvst.11.6.19.
Classification of diabetic retinopathy (DR) is traditionally based on severity grading, given by the most advanced lesion, but potentially leaving out relevant information for risk stratification. In this study, we aimed to develop a deep learning model able to individually segment seven different DR-lesions, in order to test if this would improve a subsequently developed classification model.
First, manual segmentation of 34,075 different DR-lesions was used to construct a segmentation model, with performance subsequently compared to another retinal specialist. Second, we constructed a 5-step classification model using a data set of 31,325 expert-annotated retinal 6-field images and evaluated if performance was improved with the integration of presegmentation given by the segmentation model.
The segmentation model had higher average sensitivity across all abnormalities compared to the retinal expert (0.68 and 0.62) at a comparable average F1-score (0.60 and 0.62). Model sensitivity for microaneurysms, retinal hemorrhages and intraretinal microvascular abnormalities was higher by 42.5%, 8.8%, and 67.5% and F1-scores by 15.8%, 6.5%, and 12.5%, respectively. When presegmentation was included, grading performance increased by 29.7%, 6.0%, and 4.5% for average per class accuracy, quadratic weighted kappa, and multiclass macro area under the curve, with values of 70.4%, 0.90, and 0.92, respectively.
The segmentation model matched an expert in detecting retinal abnormalities, and presegmentation substantially improved accuracy of the automated classification model.
Presegmentation may yield more accurate automated DR grading models and increase interpretability and trust in model decisions.
糖尿病视网膜病变（DR）的传统分类是基于严重程度分级，即由最严重的病变决定，但这可能会遗漏风险分层的相关信息。本研究旨在开发一种能够单独分割七种不同 DR 病变的深度学习模型，以测试其是否可以提高随后开发的分类模型的性能。
首先，使用 34075 种不同的 DR 病变的手动分割来构建分割模型，随后将其性能与另一位视网膜专家进行比较。其次，我们使用一个包含 31325 张专家标注的视网膜 6 视野图像的数据集构建了一个 5 步分类模型，并评估了在整合分割模型提供的预分割后，性能是否得到了提高。
与视网膜专家相比，分割模型在所有异常情况下的平均敏感性更高（分别为 0.68 和 0.62），且平均 F1 评分相当（分别为 0.60 和 0.62）。微动脉瘤、视网膜出血和视网膜内微血管异常的模型敏感性分别提高了 42.5%、8.8%和 67.5%，F1 评分分别提高了 15.8%、6.5%和 12.5%。当包含预分割时，平均每个类别的准确率、二次加权 kappa 和多类宏观曲线下面积的分级性能分别提高了 29.7%、6.0%和 4.5%，其值分别为 70.4%、0.90 和 0.92。
分割模型在检测视网膜异常方面与专家相当，而预分割大大提高了自动分类模型的准确性。
医学影像学
