X-Net: A Novel Deep Learning Architecture with High-resolution Feature Maps for Image Segmentation

Brain tumor segmentation is a critical task in diagnosing brain cancer and highly challenging due to the wide variety of the tumors’ location, size, and shape. Deep learning has been widely employed in the automatic segmentation of brain MRI images. There are a number of deep learning architectures proposed for seminal image segmentation with different levels of success. In general, the accuracy of the deep learning models depends on the size of the training dataset and the complexity of the model. In this paper, we proposed a novel deep learning architecture with relatively low complexity for the segmentation of brain tumors from MR images. Most of the deep learning architectures reduce the size of feature maps using techniques such as max-pooling and adding several convolutional layers. In the proposed architecture (called X-Net), the feature maps are up-sampled using transposed convolution in the first place to encode details of the input image more precisely (relying on the idea of high-resolution focusing on the details and complex structures such as camera lenses). We evaluated the proposed X-Net model using a challenging dataset for brain tumor segmentation from MR images (BraTS 2020). To verify the performance of the X-Net model, two conventional Unet models with low parameters (simple) and high parameters (complex) were also implemented. The quantitative evaluation demonstrated the superior performance of the proposed model with less trainable parameters. The X-Net model (with 0.63 M parameters) achieved a Dice score of 0.72 compared to the simple (0.54 M parameters) and complex (33.54 M parameters) Unet models with Dice indices of 0.50 and 0.67, respectively. The proposed model could be employed for the segmentation tasks with high accuracy.