Keywords:
Bones, Extremities, Conventional radiography, Digital radiography, Neural networks, Computer Applications-General, Education, Education and training
Authors:
A. Matsushima1, C. Tai-Been2, T. Okamoto1, S.-Y. Hsu2, J. Ryu-I2, N. Itayama1, T. Ishibashi1, K. Fukuda1; 1Tokyo/JP, 2Kaohsiung City/TW
DOI:
10.26044/ecr2021/C-10411
Purpose
The Japan diagnostic reference levels (DRLs) published in 2020 are based on the results of a fact-finding survey [1]. However, the DRLs include exposure doses based mainly on surface absorbed doses, which do not take into account re-exposed elements that are not recorded. The focus of this study is the lateral radiograph of the knee joint, which has a simple evaluation criterion. However, perfect positioning of the joint is not easy and multiple exposures may need to be performed before the correct position is obtained. The re-exposure rate for the lateral view of the knee is 38.2%, which is much higher than the average for radiography of other body parts.
The number of re-exposures should be minimized to reduce the medical radiation dose. We propose a support system based on a convolutional neural network (CNN) that informs the radiographer of the pattern for correction after establishing that re-exposure is require. Figure 1 shows VGG16 network as an example of a direct network. The CNN comprises a convolution layer, a rectified linear unit (ReLU), and a pooling layer. Several stacked layers (blocks) are directly learned to become the optimal mapping based on these three processes.
However, previous networks such as VGG and AlexNet, are known to overfit, and their accuracy is reduced when training a deeper network. The shortcut connection shown in Fig. 2 used information from earlier layers to solve this problem. It has been confirmed that this method, termed the residual network, prevents overfitting even when the layers are deep [2]. In this study, we used and compared a CNN model with these plain networks and residual networks.