A Comparative Analysis of Standard and Lightweight 3D U-Net and 3D DeepLabV3+ for Brain Tumor

Automated 3D segmentation of brain tumors from MRI is crucial for clinical practice. This work provides a statistically-validated comparison of four 3D deep learning architectures: standard and lightweight 3D U-Net and 3D DeepLabV3+. While standard 3D U-Net achieved the highest Whole Tumor Dice (0.884), standard 3D DeepLabV3+ demonstrated statistically significant superiority for the Tumor Core (Dice: 0.846) and Enhancing Tumor (Dice: 0.788). Critically, the lightweight models achieved a 68-72% parameter reduction and 37-41% faster inference with only a ~1.5-1.8% drop in Dice, presenting a compelling trade-off for computationally constrained environments.