An Efficient Deep Learning Approach based on 3D Res-UNet for Multimodal Brain Tumor Segmentation

Abstract

Accurate segmentation of brain tumors in MRI is an important aspect of accurate clinical diagnostics and sound surgical planning. Tumor boundary accuracy is fundamental to informing assessment of a patient’s condition, acquired through continuous expansion of Deep Learning based segmentation approaches. This study introduced an effective deep learning method to perform 3D segmentation of multimodal MRI images by enhancing the Res-UNet architecture. Our proposed model, 3D ASPP-ResUNet, incorporates an ASPP (Atrous Spatial Pyramid Pooling) module to better exploit multi-spatial scale features. The BraTS 2020 dataset has been used for training and evaluation. This model performs well according to the dice metric for different tumor regions attaining Dice scores of 0.7442 for TC (tumor core), 0.7293 for ET (enhancing tumor) and 0.8215 for WT (whole tumor). Furthermore, we observed that the 3D ASPP-ResUNet was better than currently leading models with respect to segmentation performance metrics that we defined as the Dice coefficients.