Hierarchical multi-resolution deep encoder-decoder network for MRI Brain Tumor segmentation

So far, many multi-scale encoder-decoder image segmentation architectures have been introduced to utilize the information inherited in different resolutions. However, the way the segmentation problem is formulated and the approach used for using multi-resolution vary greatly between these studies. In this paper, we first categorize and formulate the concept of multi-resolution in the deep encoder-decoder segmentation networks. Second, we found that they have the following shortcomings; (i) They need the full-resolution input images, making them unsuitable for deployment in the computational or space constrained settings; (ii) They usually designed their networks to solve the segmentation problem in one inference step; while solving complex problems in one shot may not be efficient, especially for difficult segmentation tasks. Third, based on these limitations, we introduce the Hierarchical Multi-Resolution deep encoder-decoder Networks (HMRNets) which are hierarchically learned from low-resolution to high-resolution. In this setting, the deep architecture is decomposed into a set of lower-resolution networks and the training process performed in easy-to-hard manner. As a result, HMRNets not only permits to segment the lower-resolution version of input images in case of limiting memory/hardware or communication channel capacity, but also has the potential to hierarchically use multi-resolution information, increasing the discriminative performance. Fourth, we designed a novel lightweight attentive U-shaped network named LAUNet as HMRNet baseline. Experimental analysis on the Decathlon, BraTS18 and BraTS20 brain tumor segmentation datasets demonstrate that employing LAUNet as HMRNet baseline achieves competitive performance among the new deep segmentation models.