Adaptive Pneumonia Detection Algorithm based on Deep Learning

Abstract

General convolutional neural networks are unable to automatically adjust their receptive fields for the detection of pneumonia lesion regions. This study, therefore, proposes a pneumonia detection algorithm with automatic receptive field adjustment. This algorithm is a modified form of RetinaNet with selective kernel convolution incorporated into the feature extraction network ResNet. The resulting SK-ResNet automatically adjusts the size of the receptive field. The convolutional neural network can then generate prediction bounds with sizes corresponding to those of the targets. In addition, the authors aggregated the detection results with SK-ResNet50 and SK-ResNet152 for the feature extraction network to further enhancing average precision (AP). With a data set provided by the Radiological Society of North America, the proposed algorithm with SK-ResNet50 as the feature extraction network resulted in AP 50 that was 1.5% higher than that returned by RetinaNet. The number of images processed per second differed by only 0.45, which indicated that AP was increased while detection speed was maintained. After the detection results with the SK-ResNet50 and SK-ResNet152 as the feature extraction network were combined, AP 50 increased by 3.3% compared to the RetinaNet algorithm. The experimental results show that the proposed algorithm is effective at automatically adjusting the size of the receptive field based on the size of the target, as well as increasing AP with minimal reduction in speed.