RefineFormer3D: Efficient 3D Medical Image Segmentation via Adaptive Multi-Scale Transformer with Cross Attention Fusion
The lightweight model achieves 93.44% Dice scores while running 8.35 ms per volume on GPU.
Deep Dive
Researchers Kavyansh Tyagi, Vishwas Rathi, and Puneet Goyal developed RefineFormer3D, a new 3D medical image segmentation model. It uses a novel adaptive multi-scale transformer with cross-attention fusion. With only 2.94M parameters, it achieves state-of-the-art 93.44% and 85.9% Dice scores on ACDC and BraTS benchmarks. It performs fast inference (8.35 ms/volume) with low memory use, enabling deployment in resource-constrained clinical environments for analyzing MRI and CT scans.
Why It Matters
This makes AI-powered 3D medical image analysis practical for hospitals with limited computational resources, speeding up diagnostics.