3D Transformer Models for Volumetric Medical Image Segmentation

Segmentation of volumetric medical images is
important for computer-aided diagnosis, treatment planning, and
monitoring for disease progression. However, current approaches
using CNNs may struggle to capture long-range dependencies
across three-dimensional position space, and they may lack fine

anatomical detail due to down-sampling techniques. A TransVol-
Net, a 3D transformer-based segmentation framework has been

introduced with the following components: 3D patch embedding,
a hybrid convolution – transformer encoder-decoder backbone,
and a multi-scale fusion refinement head. The model architecture
uses window-based multi-head self-attention for fast global
context modeling while leveraging convolution layers to maintain
local texture information. TransVol-Net is evaluated on the
BraTS dataset, where it achieves mean Dice scores of ≥91.0%
(WT), ≥88.0% (TC), and ≥85.0% (ET) across all the tumors, and
exceeds other, state-of-the-art methods for 3D U-Net, TransBTS,
or Swin UNTR. The results further evidence increased sensitivity
for small lesions and more fluid boundary delineation for tumor
voxels compared with other state-of-the-art segmentation models.
In conclusion, our findings for TransVol-Net demonstrate a
reformulated model that provides a more scalable and clinically
acceptable avenue for volumetric segmentation that has
applicability to CT, MRI, and PET-CT imaging workflows.
Keywords: 3D Transformer, Volumetric Segmentation, Hybrid
Encoder-Decoder, Medical Image Analysis, Multi-Scale Feature
Fusion, Deep Learning, Brain Tumor Segmentation, Dice Score