|
목적 : Design a novel deep learning neural network architecture and evaluate its performance on diabetic retinopathy (DR) diagnosis
방법 : A novel convolutional hybrid neural network named HydraNet was developed. The neural network was trained with 3127 training data (normal + DR). To evaluate diagnostic performance, another 426 test data set (217 normal, 37 mild DR, 55 moderate DR, 117 severe DR) was prepared. The mean accuracy, area under receiver operating characteristic curve (AUROC), specificity, and sensitivity were calculated. NVIDIA 1080 Ti 12GB ram GPU, Intel 8700 3.7GHz, 32 Gb ram computer was used to training and test.
결과 : he mean time to make decision was 0.042 seconds. False positive ratio was 0.921% and false negative ratio was 8.133%. Of the 217 normal subjects, 2 subjects (0.92 %) were diagnosed as mild DR and of the 209 DR patients, 17 patients (8.13 %) were diagnosed as normal. In mild DR patients, 14 patients (37%) were diagnosed as normal. In moderate DR patients, 2 patients (3.63%) were diagnosed as normal. In severe DR patients, only 1 patient (0.85 %) was diagnosed as normal. Overall AUROC was 0.956 and sensitivity / specificity was 91.8 / 98.6.
결론 : The HydraNet showed a excellent diagnostic performance and required very short time to make decision. Its performance was nearly as good as Google’s Inception V3 which was trained with about 40 times more data than HydraNet.
* This research was supported by the Bio & Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (No. NRF-2018M3A9E8066253)
|