Study on the optimal ROI cropping condition for each liver tumor type to improve the sensitivity of convolutional neural network for liver tumor ultrasound image classification

Abstract In ultrasound imaging diagnosis, not only information inside a tumor but also information around the tumor is important. Therefore, in this study, we investigated how much information about the area surrounding the tumor should be included in the region of interest (ROI) image when classifying ultrasound images of liver tumors using a convolutional neural network. Since sensitivity is important in diagnosis, we evaluated the accuracy for each type of liver tumor. We used the parameter D / L , defined as the maximum diameter of the tumor divided by the size of the ROI. As a result, the sensitivity was highest when D / L was 0.3, and the specificity was highest when D / L was 0.6 and 0.7. Therefore, to increase sensitivity, it is advisable to crop the ROI to include many surrounding areas of the liver tumor. In addition, sensitivity can be further improved by using ROI images cropped under different conditions.