Optimizing AlexNet using Swarm Intelligence for Cervical Cancer Classification

In this study, we optimized a convolutional neural network model i.e. AlexNet to classify images of cervical cancer cells. Although having canonical CNN architecture, AlexNet is only equipped with few hidden layers and thus makes it less efficient for complex objects such as cervical images. To overcome this limitation, we optimized AlexNet using a swarm-based approach (particle swarm optimization). The dataset used is the Intel & MobileODT Cervical Cancer Screening dataset. Firstly, we optimize standard AlexNet based on epoch, data subsets during training (minibatch), learning rate, input image resolution, and training-testing ratio. After having the best parameter values, we derive 3 models of AlexNet based on the number of convolutional layers. Using this approach, AlexNet with a double convolutional layer produces 60.14%, almost as good as the standard residual network on cervical images. However, when AlexNet optimized by swarm-based intelligence (particle swarm optimization) and an additional dropout layer, the accuracy can attain about 67% which is can surpass the standard residual network by 6.22%.