Tumor segmentation using CNN for automatic diagnosis of bone tumor in X-ray image.

Bone tumors are often firstly diagnosed at the nearby hospital, because they cause bone pain and swelling. The number of patients are quite small (0.3% of population), and the identification of malignancy requires a great effort even for specialists. Therefore, there is a need to develop a system that can automatically extract bone tumor regions and estimate benign and malignant conditions. In this study, we propose a method for automatically extracting bone tumor regions using convolutional neural networks. We compared the segmentation performance of U-Net, DeepLab v3, and Mask R-CNN, and inspected the effect of tumor area cropping before semantic segmentation. Experiments using 118 knee X-ray images of 18 patients revealed that U-Net achieved the highest dice coefficient (0.67) in original x-ray image. In the experiment using cropped x-ray images, DeepLab v3 achieved highest 0.92 dice coefficient.