|
KMID : 0390320180280020029
|
|
Chungbuk Medical Journal
2018 Volume.28 No. 2 p.29 ~ p.36
|
|
|
Grading diabetic macular edema of retinal images using transfer learning with deep neural network
|
|
Lee Tae-Soo
|
|
|
Abstract
|
|
|
Purpose: This paper proposed a method of grading diabetic macular edema(DME) of retinal images using transfer learning with inception v3, which is pre-trained deep neural network(DNN) by millions of non-medical images, and has 48 layers with 23 million parameters, and uses 299x299 images as DNN input data.
Materials and methods: The data set for training and testing the DNN was Messidor database publicly opened with 1200 retinal images for the studies on computer-assisted diagnoses of diabetic retinopathy. The data was divided by 7:3 ratio for training and testing and was pre-processed to normalize the size, and to reduce the impacts with various image acquisition conditions by three step.
Results: After 50 epochs of deep learning with initial learning rate 0.001, which was reduced by a factor of 0.5 every 10 epochs, the DNN acquired 91.64% accuracy for grading the DME of 359 testing retinal images.
Conclusion: Therefore, it is expected to help improve the ophthalmologist’s diagnostic accuracy of DME with the objective and quantitative grading information provided by the proposed DNN.
|
|
|
KEYWORD
|
|
|
deep neural network, diabetic macular edema, retinal image, transfer learning
|
|
|
FullTexts / Linksout information
|
|
|
|
|
|
Listed journal information
|
|
|
|