Automatic brain tumor segmentation from Multiparametric MRI based on cascaded 3D U-Net and 3D U-Net++

Purpose: Brain tumor is often a deadly disease and its diagnosis and treatment are challenging tasks for physicians for the heterogeneous nature of the tumor cells. Automatic, accurate segmentation of brain tumors can be a significant tool to assist physicians in the diagnosis of brain diseases. Existing methods can achieve general results, the segmentation accuracy not comparable to that of manual segmentation by experienced physicians, especially in enhanced tumor regions. Methods: We trained cascaded 3D U-Net and 3D U-Net++ networks to realize the automatic segmentation of brain tumors in Magnetic Resonance Imaging (MRI) images from the Brain Tumor Segmentation Challenge 2019 dataset (BRATS 2019). First, we decompose the segmentation of brain tumor into the segmentation of the whole tumor (WT), tumor core (TC) and enhanced tumor (ET). Second, we train the models in axial, coronal, and sagittal plane images. We then fuse the results from the three views to produce the final segmentation result. In particular, the U-Net++ is used to segment the enhanced tumor for the latter's more complex structure compared with other sub-regions. We also tested the performance of the methods on a clinical MRI image dataset with manual standard tumor contours. Results: The networks' performances were verified on BRATS 2019 images. On the clinical dataset, we got DSC metric values of 0.890, 0.842, and 0.835 for the complete, core, and enhanced regions respectively. Segmentation performance on the clinical dataset, especially the performance of 3D-UNet++, has been approved by physicians. Conclusion: The method's performance is clinically of significance.