Purpose: This study aims to implement three-dimensional convolutional neural networks (3D-CNN) for clinical target volume (CTV) segmentation for whole breast irradiation and investigate the performance with and without shape regularization model (SRM).
Methods: The 3D-UNet CNN was adopted to conduct automatic segmentation of the CTV for breast cancer. The 3D-UNet was trained using the 192 patients’ dataset which contains both left- and right-sided breast cancer patients. SRM was introduced to the 3D-UNet model to improve the segmentation performance by modifying the segmentation shape. SRM was trained in advance using the dataset of the incomplete CTV shapes and the ground truth CTVs. In the 3D-UNet+SRM model, the customized loss function was employed. The segmentation accuracy was evaluated using the Dice similarity coefficient (DSC) and Hausdorff distance (HD). To compare the performance of the 3D-UNet alone and 3D-UNet+SRM, the Wilcoxon signed-rank test was applied to the DSC and HD scores of the test dataset.
Results: For the 3D-UNet alone, the DSC and the HD for the test dataset were 0.852 and 57.80, respectively. The DSC and the HD for the 3D-UNet+SRM were 0.854 and 52.67. The Wilcoxon signed-rank test demonstrated that both DSC and HD of the 3D-UNet+SRM performed better than 3D-UNet alone. The incomplete structures like small holes, jagged edges, and imperfect shapes were eliminated or compensated with SRM. SRM also worked to modify the images which had islands predicted as part of the CTV on the opposite side of the breast. However, if the 3D-UNet predicted the opposite side of the breast as the target breast, the small island on the correct side was eliminated.
Conclusion: SRM modified the incomplete CTV shapes during the training process of 3D-UNet segmentation. If we make the SRM more sophisticated with a larger dataset, 3D-UNet+SRM could be improved more.