Title: Performance of convolutional neural networks optimisers: an extensive evaluation on glaucoma prediction
Authors: Kishore Balasubramanian
Addresses: Dr. Mahalingam College of Engineering and Technology, Pollachi, 642003, India
Abstract: This paper focuses on improving the CNN performance by optimising the architecture parameters through optimisers and loss function in predicting glaucoma on locally collected retinal fundus images. Four gradient descent-based optimisers were compared, i.e., stochastic gradient descent (SGD), adaptive gradient (AdaGrad), adaptive delta (Adadelta) and adaptive momentum (Adam). Mean squared error (MSE) and binary cross-entropy (BCE) were the loss functions chosen. The dataset was divided into 60% training and 40% testing. Two CNN architectures, AlexNet, and ResNet were developed and trained on the dataset with 0.01 learning rate. The number of epochs set was 50, 100 and 200 with batch size 60. The methods were evaluated in terms of MSE and accuracy. Adam achieved the lowest training loss with appreciable accuracy and also outperforms other adaptive techniques. The assessment demonstrated that Adam-based optimisation in CNN was able to diagnose glaucoma accurately with less loss and better convergence speed.
Keywords: glaucoma; fundus image; convolutional neural networks; CNNs; deep learning; optimisation.
International Journal of Computational Intelligence Studies, 2021 Vol.10 No.4, pp.217 - 231
Received: 22 Jun 2020
Accepted: 15 Nov 2020
Published online: 20 Jan 2022 *