Network Pruning for OCT Image Classification

Convolutional neural network (CNN) has expanded rapidly, and has been widely used in medical image classification. The large number of parameters in a neural network makes CNN models computationally expensive. This leads to slow inference speed, especially for 3D data such as optical coherence tomography (OCT) for retinal images. A volume OCT scan of retina often contains hundreds of 2D images which needs to be analyzed sequentially in a local computer with limited computational resources. We introduce network pruning to OCT images classification and propose an algorithm to prune networks. We compress the popular classification models, such as ResNet and VGG. For example, within 1% accuracy loss, we compress ResNet-18 from 44.8 MB to 69 KB and VGG-16 from 537.1 MB to 194 KB. These pruned models are much smaller and easier to deploy on the OCT devices. As for the inference speed, the pruned models are 10 to 20 times faster than original models for ResNet and VGG in CPU.
© Springer Nature Switzerland AG 2019.