Optimal cropping for input images used in a convolutional neural network for ultrasonic diagnosis of liver tumors

In recent years there have been many studies on computer-aided diagnosis (CAD) using convolutional neural networks (CNNs). For CAD of a tumor, data are generally obtained by cropping a region of interest (ROI), including a tumor, in an image. However, ultrasonic diagnosis also uses information from around a tumor. Therefore, in CAD using ultrasound images, diagnostic accuracy could be improved by using a ROI that includes the periphery of the tumor. In this study, we examined how much of the surrounding area should be included in a ROI for a CNN using ultrasound images of liver tumors. We used the ratio between the maximum diameter of the tumor and the ROI size as the index for ROI cropping. Our results show that the diagnostic accuracy was maximized when this index is 0.6. Therefore, optimal ROI cropping is important in CNNs for ultrasonic diagnosis.