Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

3D UX-Net: A Large Kernel Volumetric ConvNet Modernizing Hierarchical Transformer for Medical Image Segmentation

Ho Hin Lee, Shunxing Bao, Yuankai Huo, Bennett A. Landman

TL;DR

This work addresses the high computational cost of 3D ViT-based medical image segmentation by proposing 3D UX-Net, a purely ConvNet architecture that emulates hierarchical transformer behavior using volumetric depthwise convolutions with large kernels and depthwise channel-wise scaling. The encoder leverages LK-based depthwise convolutions (starting at $7\times7\times7$) with an inverted bottleneck to widen channel representations, while a ConvNet-based decoder fuses multi-scale features through skip connections. Across FeTA2021, FLARE2021, and AMOS2022, 3D UX-Net achieves consistently higher Dice scores than SwinUNETR, with faster convergence and notable transfer-learning gains (e.g., mean Dice $0.900$ on AMOS2022, a $2.27\%$ improvement). The results suggest a practical, parameter-efficient alternative to transformer backbones for high-resolution 3D segmentation, with strong potential for clinical deployment and further efficiency enhancements.

Abstract

The recent 3D medical ViTs (e.g., SwinUNETR) achieve the state-of-the-art performances on several 3D volumetric data benchmarks, including 3D medical image segmentation. Hierarchical transformers (e.g., Swin Transformers) reintroduced several ConvNet priors and further enhanced the practical viability of adapting volumetric segmentation in 3D medical datasets. The effectiveness of hybrid approaches is largely credited to the large receptive field for non-local self-attention and the large number of model parameters. In this work, we propose a lightweight volumetric ConvNet, termed 3D UX-Net, which adapts the hierarchical transformer using ConvNet modules for robust volumetric segmentation. Specifically, we revisit volumetric depth-wise convolutions with large kernel size (e.g. starting from $7\times7\times7$) to enable the larger global receptive fields, inspired by Swin Transformer. We further substitute the multi-layer perceptron (MLP) in Swin Transformer blocks with pointwise depth convolutions and enhance model performances with fewer normalization and activation layers, thus reducing the number of model parameters. 3D UX-Net competes favorably with current SOTA transformers (e.g. SwinUNETR) using three challenging public datasets on volumetric brain and abdominal imaging: 1) MICCAI Challenge 2021 FLARE, 2) MICCAI Challenge 2021 FeTA, and 3) MICCAI Challenge 2022 AMOS. 3D UX-Net consistently outperforms SwinUNETR with improvement from 0.929 to 0.938 Dice (FLARE2021) and 0.867 to 0.874 Dice (Feta2021). We further evaluate the transfer learning capability of 3D UX-Net with AMOS2022 and demonstrates another improvement of $2.27\%$ Dice (from 0.880 to 0.900). The source code with our proposed model are available at https://github.com/MASILab/3DUX-Net.

3D UX-Net: A Large Kernel Volumetric ConvNet Modernizing Hierarchical Transformer for Medical Image Segmentation

TL;DR

This work addresses the high computational cost of 3D ViT-based medical image segmentation by proposing 3D UX-Net, a purely ConvNet architecture that emulates hierarchical transformer behavior using volumetric depthwise convolutions with large kernels and depthwise channel-wise scaling. The encoder leverages LK-based depthwise convolutions (starting at ) with an inverted bottleneck to widen channel representations, while a ConvNet-based decoder fuses multi-scale features through skip connections. Across FeTA2021, FLARE2021, and AMOS2022, 3D UX-Net achieves consistently higher Dice scores than SwinUNETR, with faster convergence and notable transfer-learning gains (e.g., mean Dice on AMOS2022, a improvement). The results suggest a practical, parameter-efficient alternative to transformer backbones for high-resolution 3D segmentation, with strong potential for clinical deployment and further efficiency enhancements.

Abstract

The recent 3D medical ViTs (e.g., SwinUNETR) achieve the state-of-the-art performances on several 3D volumetric data benchmarks, including 3D medical image segmentation. Hierarchical transformers (e.g., Swin Transformers) reintroduced several ConvNet priors and further enhanced the practical viability of adapting volumetric segmentation in 3D medical datasets. The effectiveness of hybrid approaches is largely credited to the large receptive field for non-local self-attention and the large number of model parameters. In this work, we propose a lightweight volumetric ConvNet, termed 3D UX-Net, which adapts the hierarchical transformer using ConvNet modules for robust volumetric segmentation. Specifically, we revisit volumetric depth-wise convolutions with large kernel size (e.g. starting from ) to enable the larger global receptive fields, inspired by Swin Transformer. We further substitute the multi-layer perceptron (MLP) in Swin Transformer blocks with pointwise depth convolutions and enhance model performances with fewer normalization and activation layers, thus reducing the number of model parameters. 3D UX-Net competes favorably with current SOTA transformers (e.g. SwinUNETR) using three challenging public datasets on volumetric brain and abdominal imaging: 1) MICCAI Challenge 2021 FLARE, 2) MICCAI Challenge 2021 FeTA, and 3) MICCAI Challenge 2022 AMOS. 3D UX-Net consistently outperforms SwinUNETR with improvement from 0.929 to 0.938 Dice (FLARE2021) and 0.867 to 0.874 Dice (Feta2021). We further evaluate the transfer learning capability of 3D UX-Net with AMOS2022 and demonstrates another improvement of Dice (from 0.880 to 0.900). The source code with our proposed model are available at https://github.com/MASILab/3DUX-Net.
Paper Structure (20 sections, 1 equation, 4 figures, 7 tables)

This paper contains 20 sections, 1 equation, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Transformer-based Segmentation
  4. Depthwise convolution based Segmentation
  5. 3D UX-Net: Intuition
  6. 3D UX-Net: Complete Network Description
  7. Depth-wise Convolution Encoder
  8. Decoder
  9. Experimental Setup
  10. Results
  11. Evaluation on FeTA & FLARE
  12. Transfer Learning with AMOS
  13. Ablation Analysis
  14. Discussion
  15. Conclusion
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Overview of our proposed designed convolution blocks to simulate the behaviour of swin transformers. We leverage depthwise convolution and pointwise scaling to adapt large receptive field and enrich the features through widening independent channels. We further compare different backbones of volumetric ConvNets and Swin Transformer block architecture. The yellow dotted line demonstrates the differences in spatial position of widening feature channels in the network bottleneck.
  • Figure 2: Overview of the proposed 3D UX-Net with our designed convolutional block as the encoder backbone. LK convolution is used to project features into patch-wise embeddings. A downsampling block is used in each stage to mix and enrich context across all channels, while our designed blocks extract meaningful features in depth-wise setting.
  • Figure 3: Validation Curve with Dice Score for FeTA2021 (a), FLARE2021 (b) and AMOS2022 (c). 3D UX-Net demonstrates the fastest convergence rate with limited samples training (FeTA2021) and transfer learning (AMOS2022) scenario respectively, while the convergence rate is comparable to SwinUNETR with the increase of sample size training (FLARE2021).
  • Figure 4: Qualitative representations of tissues and multi-organ segmentation across three public datasets. Boxed are further zoomed in and visualize the significant differences in segmentation quality. 3D UX-Net shows the best segmentation quality compared to the ground-truth.