Cascaded MNetV3UNet: A Lightweight Two-Stage   Architecture for High-Precision Brain Tumor Segmentation in MRI

Precise brain tumor segmentation is an essential but quite challenging process in MRI images because of their irregular shapes, heterogeneous appearance, and low contrast with surrounding tissues. While U-Net-based architectures have achieved significant success, their high computational complexity limits deployment on resource-constrained systems. In this study, a Novel two-stage cascaded architecture, MNetV3UNet, is introduced, which employs the light-weight MobileNetV3-Large as an encoder and the standard U-Net as the decoder. MobileNetV3 block consists of a sequence of blocks that generate feature maps enhanced using Inverted Residual Blocks and squeeze-and-excitation modules. The Unet decoder consists of an iterative process of upsampling, interpolation, concatenation, and refinement. The first stage produces a coarse segmentation, which is refined in the second stage to enhance boundary accuracy and detail. The cascaded approach leverages a multi-scale feature extraction process, which is also coupled with the progressive refinement method. This way, we ensure a much higher degree of segmentation precision while still maintaining computational efficiency. The proposed model has been tested on the BraTS 2020 dataset, yielding a Dice score of 88.35 for Whole Tumor (WT), 89.03 for Tumor Core (TC), and 92.30% for Enhancing Tumor (ET). Additionally, it achieved a Jaccard score of 83.76 for WT, 86.15% for TC, and 89.68% for ET. The specificity obtained for WT was 99.72, for TC, 99.89, and for ET, 97.96%. It achieved sensitivity of 88.91 for WT, 89.37 for TC, and 92.88% for ET.  These outcomes provide clear evidence of the proposed innovative architecture's ability to achieve an excellent balance between segmentation accuracy and computational efficiency.