Multi-parametric MRI scans are used to accurately identify the morphological variability of glioblastoma sub-regions of the brain tumor. For this purpose, T1, T1CE, T2, and FLAIR modalities are used to identify essential information of about ET, TC and WT for treatment planning. Small ET regions struggle to be identified by traditional architecture because the area occupied by ET is less than 2% of brain volume, and even spatial resolution adoption is one of the challenges. The paper proposes an Enhanced Neural Architecture Search-based Feature Pyramid Network UNet3D (NAS-FPN-UNet3D) that incorporates autonomous operation selection inside multi-scale feature pyramids. In NAS-FPN-UNet3D, adaptive feature reconfiguration is managed by Convolutional Block Attention Module (CBAM). To maintain image quality and spatial accuracy, gradient stability is managed through residual blocks. NAS is one of the important components for effectively refining long-range information through five parallel operations: Standard Convolution, Dilated Convolution, Depth-wise Separable Convolution (with 94.7% parameter reduction), Asymmetric Convolution and identify connection. Operations are automatically selected based on the learnable parameters; for contextual modelling, dilated convolutions are preferred at coarse levels, while for efficiency at finer levels, depth-wise operations are employed. When evaluated on the BraTS2020 dataset, the proposed model uses parameters and yields Dice scores of 0.8406 ± 0.05 (ET), 0.868 ± 0.06 (TC), and 0.9157 ± 0.03 (WT) with corresponding Hausdorff distances of 4.57mm ± 0.05 mm, 6.75 mm ± 0.04 mm, and 6.85 mm ± 0.02mm, respectively. The proposed model requires 34% fewer parameter than fixed-convolution baselines while achieving greater accuracy.