-blue){kind=icon}
This repository provides the official codebase for the paper:
Optimized High Resolution 3D Dense-U-Net Network for Brain and Spine Segmentation
published in MDPI Applied Sciences, 2019.
If you are working on medical 3D segmentation and seek to benchmark against our methods, this repository offers the corresponding academic implementation.
- Optimized Dense-U-Net architecture for brain MRI and spine CT segmentation
- Improvements over standard U-Net in segmentation accuracy
- Designed for practical training on available GPUs in 2019
- Implementations include:
- 2D and 3D Dense-U-Net
- 2D and 3D Res-U-Net
- Classic U-Net for baseline comparison
- Spine Dataset: Previously available through SpineWeb. Offline as of April 2025
- Brain MRI Dataset: Private; redistribution is unfortunately not permitted.
Segmentation examples using 3D Dense-U-Net:
| Brain MRI Slice | Dense-U-Net Architecture | Brain Segmentation Output | Spine Segmentation Output |
|---|---|---|---|
- Figure 1: Example of MRI brain slice and thoracic CT scan. Segmented tissue is highlighted.
- Figure 2: Dense-U-Net architecture (green: residual connections, blue: dense connections).
- Figure 3: 3D brain segmentation output.
- Figure 4: 3D spine segmentation output (vertebrae abnormalities included in ground truth).
If you reference or build on this work, please cite:
@article{kolarik2019optimized,
title={Optimized High Resolution 3D Dense-U-Net Network for Brain and Spine Segmentation},
author={Kola{\v{r}}{\'\i}k, Martin and Burget, Radim and Uher, V{\'a}clav and {\v{R}}{\'\i}ha, Kamil and Dutta, Malay Kishore},
journal={Applied Sciences},
volume={9},
number={3},
pages={404},
year={2019},
publisher={Multidisciplinary Digital Publishing Institute}
}
This implementation is inspired by:
- Deep Learning Tutorial for Kaggle Ultrasound Nerve Segmentation
- The One Hundred Layers Tiramisu (Fully Convolutional DenseNets)
- Densely Connected Convolutional Networks (DenseNet)
- U-Net: Convolutional Networks for Biomedical Image Segmentation
This repository is no longer actively maintained and will be archived.
The code serves as an academic reference for reproducing results from the original publication.
For current state-of-the-art (SOTA) research, please consult newer architectures and frameworks.