ANALYSIS OF TRANSFER LEARNING FOR SELECT RETINAL DISEASE CLASSIFICATION.
ANALYSIS OF TRANSFER LEARNING FOR SELECT RETINAL DISEASE CLASSIFICATION.
复制标题
DOI：
10.1097/iae.0000000000003282
复制
发表时间：
2022-01-01
期刊：
影响因子：
--
通讯作者：
Fernandez-Granda C
中科院分区：
文献类型：
--
作者：
Gelman R;Fernandez-Granda C
分析迁移学习（TL）在眼底照相分类糖尿病视网膜病变（DR）和光谱域光学相干断层扫描（SD-OCT）筛选视网膜病变中的作用。五个广泛使用的开源深度神经网络（DNN）和四个定制的更简单和更小的网络（称为CBR系列）在两个任务上进行了训练和评估：（1）使用眼底照相对DR进行分类，以及（2）使用SD-OCT对玻璃疣、脉络膜新生血管（CNV）和糖尿病性黄斑水肿（DME）进行分类。二次加权Kappa系数用于测量每个网络和地面真实标记的测试用例之间的一致性水平。对于基于SD-OCT的分类，计算每个网络的准确度。Kappa和准确性进行了比较，使用和不使用TL的每个网络的迭代，以评估其效果。对于DR分类，Kappa随所有网络的TL增加而增加（增加范围为0.152-0.556）。对于基于SD-OCT的分类，5个开源DNN中有4个的准确性增加（范围增加1.8-3.5%），其余DNN略有下降（-0.6%），4个CBR网络中有3个略有下降（范围减少0.9-1.8%），其余CBR网络下降了9.6%。TL改善了所有网络的DR分类的性能，由Kappa测量，尽管效果从小到大不等。TL对所分析的9个网络中的8个网络的基于SD-OCT的分类的准确性影响极小。这些结果意味着TL可以显著提高DR分类的性能，但对基于SD-OCT的分类的影响可能很小。
To analyze the effect of transfer learning (TL) for classification of diabetic retinopathy (DR) by fundus photography and select retinal diseases by spectral-domain optical coherence tomography (SD-OCT). Five widely used open-source deep neural networks (DNNs) and 4 customized simpler and smaller networks, termed the CBR-family, were trained and evaluated on 2 tasks: (1) classification of DR using fundus photography and (2) classification of drusen, choroidal neovascularization (CNV), and diabetic macular edema (DME) using SD-OCT. For DR-classification, the quadratic weighted Kappa coefficient was used to measure the level of agreement between each network and ground-truth labeled test cases. For SD-OCT based classification, accuracy was calculated for each network. Kappa and accuracy were compared between iterations with and without use of TL for each network to assess for its effect. For DR-classification, Kappa increased with TL for all networks (range of increase 0.152–0.556). For SD-OCT based classification, accuracy increased for 4 of 5 open-source DNNs (range of increase 1.8–3.5%), slightly decreased for the remaining DNN (−0.6%), decreased slightly for 3 of 4 CBR networks (range of decrease 0.9–1.8%) and decreased by 9.6% for the remaining CBR network. TL improved performance, as measured by Kappa, for DR-classification for all networks, although the effect ranged from small to substantial. TL had minimal effect on accuracy for SD-OCT based classification for 8 of the 9 networks analyzed. These results imply that TL may substantially increase performance for DR-classification but may have minimal effect for SD-OCT based classification.