Open AccessDeep convolutional neural networks for the detection of macular diseases from optical coherence tomography images
Author(s) -
Iqbal Kharisudin,
Mien Fatimah Az-zahra,
Endang Retno Winarti,
St. Budi Waluya
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1567/2/022076
Subject(s) - convolutional neural network , optical coherence tomography , computer science , epoch (astronomy) , deep learning , artificial neural network , artificial intelligence , stride , variation (astronomy) , pattern recognition (psychology) , algorithm , computer vision , ophthalmology , medicine , physics , astrophysics , stars , computer security
The purpose of this research is to design a system to recognize CNV image, DME, Drusen, and Normal, which were produced by Optical Coherence Tomography (OCT). A system would provide a training model, evaluation, and accuracy value. We used the Convolution Neural Network method with default parameter 50 epoch, one stride, 83484 train data images, and learning rate value 0.001 with the help of Python 3.7 software. The examination using epoch variation, stride, number of train data, and learning rate value resulted in different accuracy values. According to epoch variation, the best accuracy was 50 epoch with an accuracy value of 0.99 and loss validation of 0.2034. The best accuracy of stride value variation was one stride with an accuracy value of 0.99 and loss validation of 0.2267. The best accuracy of train data variation was 83484 images with an accuracy value of 0.99 and a loss validation value of 0.2524. The learning rate variation value with the best accuracy was 0.0001, with an accuracy value of 0.992 and validation loss value of 0.2524. According to the result of the research, it was obtained that convolution neural network architecture gain the best model with accuracy value 0.992 according to variation parameter 50 epoch, one stride, 83484 train data images, and learning rate 0.0001.