DALight-3D: A Lightweight 3D U-Net for Brain Tumor Segmentation from Multi-Modal MRI
30% fewer parameters, yet outperforms Residual 3D U-Net on MRI benchmark.
Automatic brain tumor segmentation from multi-modal MRI is critical for diagnosis and treatment planning, but existing 3D deep learning models are often computationally heavy. A new paper from researchers at (affiliation not specified in abstract) introduces DALight-3D, a lightweight 3D U-Net variant designed to reduce parameters while maintaining segmentation accuracy. The model combines four key innovations: depthwise separable 3D convolutions (SepConv) to cut FLOPs, identifier-conditioned normalization to handle multi-modal inputs, cross-slice attention (CSA) to capture volumetric context, and adaptive skip fusion (SSFB) to blend low- and high-level features.
Evaluated on the Medical Segmentation Decathlon Task01 BrainTumour benchmark, DALight-3D achieves a mean Dice coefficient of 0.727 with just 2.22 million parameters. This compares favorably against standard 3D U-Net (0.710 Dice, 3.20M params) and other baselines like Attention U-Net and V-Net. Ablation studies confirm that each component contributes to performance—removing SepConv, CSA, or SSFB leads to consistent degradation. The results demonstrate that DALight-3D offers a practical, resource-efficient solution for deployment in clinical settings where compute or memory is constrained.
- DALight-3D uses depthwise separable 3D convolutions, cross-slice attention, and adaptive skip fusion to shrink model size.
- Achieves 0.727 mean Dice on Medical Segmentation Decathlon Task01 with only 2.22M parameters—30% fewer than Residual 3D U-Net.
- Ablation experiments confirm every component (SepConv, identifier normalization, CSA, SSFB) is essential for top performance.
Why It Matters
Enables accurate brain tumor segmentation on less powerful hardware, potentially broadening access to AI-assisted radiology.