<![CDATA[One of the deadliest diseases worldwide is brain tumors. In identifying brain tumors, experts perform a subjective analysis that requires considerable time. Previous research has developed automatic 3D brain tumor segmentation using Deep Learning (DL) approaches such as 3D UNet and 3D ResNet. However, these approaches demand significant computational resources. In resource-constrained settings, key criteria for determining the best architecture include memory consumption, inference speed, and accuracy. Therefore, this study introduces the development of the 3D S-UNet architecture, constructed by combining 3D ShuffleNet-V2 as an encoder and 3D UNet as a decoder. The integration of these 3D data processors allows the architecture to be more precise in identifying brain tumor locations and capture richer feature values compared to 2D data processing. The researchers compare 3D S-UNet with another Lightweight Deep Learning architecture, 3D Mobile-UNet. The results show that 3D S-UNet has a smaller memory consumption, using 0.56GB for the highest allocated memory and 1.71GB for reserved memory. In terms of inference speed, 3D S-UNet is faster compared to the other three architectures, achieving a speed of 135.881 milliseconds. 3D S-UNet demonstrates favorable results with a Whole Tumor (WT) dice score, sensitivity, and specificity of 83%, 85%, and 88%, respectively.]]>
Copyrights © 2025
