A COMPACT MOBILENETV3-UNET FOR RESOURCE-EFFICIENT BRAIN TUMOR SEGMENTATION

Abstract

Introduction: Manual segmentation of brain tumors is labor-intensive and susceptible to inter-observer variability, yet accurate segmentation using magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, radiation therapy, and follow-up monitoring. Current approaches focus on complex 3D networks, transformers, or ensembles of multiple models, all requiring extensive GPU memory and training time, making them impractical for resource-constrained environments.

Objective: This paper presents a lightweight segmentation framework based on the MobileNetV3-UNet architecture.

Methodology: The model was trained and evaluated on BraTS2021, with initial validation on BraTS2020, segmenting the Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) across four MRI modalities (FLAIR, T1, T1CE, T2). To reduce reliance on costly expert annotation, pseudo-label semi-supervised learning (SSL) was evaluated under limited-label budgets (20% and 40%).

Results: The fully supervised model achieved a Mean DSC of 0.9127 on BraTS2021 (WT: 0.9385, TC: 0.9114, ET: 0.8882) with only 10.1 M parameters and 2.14 GFLOPs, substantially reducing computational cost compared to heavy state-of-the-art architectures. Under limited labels, patient-wise 3D evaluation showed consistent improvement, with Mean DSC increasing from 0.8862 (supervised 20%) to 0.8919 (SSL 20%) and 0.8942 (SSL 40%).

Conclusion: These results demonstrate that a lightweight network can achieve competitive segmentation accuracy at a fraction of the computational cost of heavier models, and that unlabeled data can enhance performance when annotated data is scarce.