An Adaptive Segmentation Technique to Detect Brain Tumors Using 2D Unet

The UNet is one of the most well-known convolutional neural network (CNN) architectures used for biomedical image segmentation. Unfortunately, the 2D variant is typically discouraged for volumetric brain tumor segmentation due to slices being correlated with one another. Thus, 3D-Unets have become prevalent in the annual Multimodal Brain tumor Segmentation Challenge (BRaTS) hosted by the Perelman School of Medicine of University of Pennsylvania (UPenn). However, with unique data preprocessing and generator techniques, 2D-Unets may achieve competitive accuracy and performance with 3D-Unets. Furthermore, the addition of residual blocks (R) and squeeze-and-excitation (SE) blocks in the upsampling portion of 2D-Unets could further speed up performance and minimize computational costs without sacrificing f1 or Jaccard's similarity score. This reveals that 2D-UNets for 3D biomedical image segmentation are still valuable. This paper involves the detailed comparison between 2D-Unets and 2D-SE-RUNets for the purposes of segmenting a whole high-grade glioma (HGG) using the metrics of Jaccard's similarity, recall, specificity, and precision. Results indicate that the 2D-SE-RUNet model is superior to the traditional 2D UNet due to efficieny, which can benefit those looking to save computational costs and time.