Segmentation of renal tumors in CT images by 3D U-Net preserving rotational symmetry in axial slices
Ryo Tanimoto, Kokoro Higuchi, Toshitaka Ishiguro, Kensaku Mori, Kosuke Kojo, Takahiro Kojima, Hideki Kakeya · Optics Continuum · 2022
We present a method to improve accuracy of renal tumor segmentation in CT images by considering the structure of kidneys. We apply deep learning based on the following protocol. First, kidney regions are extracted from the CT images. Second, adjacent eight slices along axial direction are picked up as a patch. Third, the center of gravity of the kidney region in the patch is aligned to the center of the patch in sagittal and coronal direction. Fourth, we apply data augmentation with scaling and rotation around the center, which preserves the basic slice structure of kidneys. Finally, these patches are fed to a 3D U-Net for training. Compared with the conventional 3D U-Net, the proposed method improves the DICE score from 0.507 to 0.604.