Performance evaluation of mask R-CNN for lung segmentation using computed tomographic images

Image segmentation techniques based on machine learning are able to improve diagnostic and therapeutic accuracy by localizing target areas. The accuracy and efficiency of these techniques are dependent on network architecture and loss minimization method because the performance of a machine learning model is determined by training strategies. In this study, the lung segmentation based on computed tomographic images was performed by using mask regional convolutional neural networks (R-CNNs) with various feature extraction networks and optimizers. The effects of the feature extraction networks and optimizers on the trained mask R-CNNs were evaluated in terms of total training loss, segmentation accuracy and training time. The results showed that the convergence of total loss values during network training was affected by the architectures of the feature extraction networks as well as the optimizers. The lung segmentation accuracy and training time of the mask R-CNN were mainly dependent on the optimizer and network architecture, respectively. Among the various optimizers, the ASGD optimizer maximized lung segmentation accuracy, and the training time was reduced by the feature extraction network including general convolution layers and feature pyramid network (FPN). In conclusion, it is important to apply the optimal network architecture and optimizer to the mask R-CNN for maximizing its performance, and the optimized mask R-CNN can be potentially used for improving diagnostic and therapeutic accuracy.