3D Brain Tumor Segmentation on the HP ZGX Nano AI Station

Customer Case Study | Christian Chabrerie, Ph.D École pour l’informatique et les Techniques Avancées (EPITA)

Accelerating Medical AI Development with On-Premises GPU Computing

Brain tumor segmentation from MRI scans is one of the most computationally demanding tasks in medical AI. The MICCAI BraTS Challenge (http://braintumorsegmentation.org/) provides a standardized benchmark for this problem, requiring models to delineate three distinct tumor sub-regions from multimodal 3D MRI volumes: the necrotic tumor core, peritumoral edema, and enhancing tumor tissue.

For AI developers working on medical imaging such as this, a key challenge is establishing an effective development workflow. Training and inference on 3D volumetric data require substantial GPU memory and compute capacity, particularly when working with sensitive healthcare datasets and iterative model development. Researchers needed a local AI development environment capable of handling large 3D volumetric models with sensitive patient data, enabling rapid iteration without reliance on external compute infrastructure.

My students’ teams at EPITA developed a Residual 3D U-Net approach with a specialized localization technique that centers and reduces input image size to focus on the tumor region. This architectural choice enables more accurate predictions while reducing computational requirements but still demands serious GPU resources for both training and inference.

The HP ZGX Nano AI Station, powered by the NVIDIA GB10 Grace Blackwell Superchip1, provides a development environment perfectly suited for compute-intensive medical AI workloads. The system’s 128GB unified LPDDR5x memory reduces memory bottlenecks that typically constrain 3D medical imaging models, while the Blackwell architecture delivers the compute throughput needed in a compact desktop form factor.

Running NVIDIA DGX™ OS, the ZGX Nano provides a pre‑configured PyTorch development environment ready for use. The implementation uses FLAIR and T1Ce MRI modalities as inputs, balancing segmentation accuracy with computational efficiency. The model processes NIfTI volumetric files, performs inference across the full 3D volume, and generates both 2D slice visualizations and 3D surface reconstructions of detected tumor regions.

EPITA - School of Engineering & Computer Science 2

The Challenge

The Solution

“My students’ teams at EPITA developed a Residual 3D U-Net approach with a specialized localization technique that centers and reduces input image size to focus on the tumor region. This architectural choice demands serious GPU resources for both training and inference.

The HP ZGX Nano AI Station, powered by the NVIDIA GB10 Grace Blackwell Superchip1, provides a development environment perfectly suited for compute- intensive medical AI workloads.” Christian Chabrerie, École pour l’informatique et les Techniques Avancées (EPITA)

http://braintumorsegmentation.org/

The implementation leverages a Residual 3D U-Net architecture with a specialized localization preprocessing step. This technique centers the input volume on the tumor region and reduces spatial dimensions, enabling the model to focus computational resources on relevant anatomy while maintaining segmentation precision.

• �Dual-modality input (FLAIR + T1Ce) balances diagnostic information with compute efficiency

• �Cross-validation during training improves generalization across patient populations

• �Minimum pixel thresholding in post-processing reduces false positive predictions

• �Label remapping creates contiguous class indices [0-3] for standard loss functions

Technical Implementation

Key Architectural Decisions

The segmentation model produces both 2D slice overlays and 3D volumetric reconstructions. The color mapping identifies tumor sub-regions: green indicates peritumoral edema, blue shows the necrotic/non-enhancing tumor core, and red highlights the enhancing tumor, the most aggressive component characterized by blood-brain barrier breakdown.

Model Output Visualization

EPITA - School of Engineering & Computer Science 3

Prediction Overlay: Ground Truth Overlay:

Prediction 3D: Ground Truth 3D:

2D Slice Comparison: Prediction vs. Ground Truth

3D Tumor Surface Reconstruction: Prediction vs. Ground Truth

EPITA - School of Engineering & Computer Science 4

“The color mapping identifies tumor sub-regions: green indicates peritumoral edema, blue shows the necrotic/non-enhancing tumor core, and red highlights the enhancing tumor, the most aggressive component characterized by blood-brain barrier breakdown.”

Christian Chabrerie, École pour l’informatique et les Techniques Avancées (EPITA)

The complete implementation is available as open source, including model architecture, training scripts, and the Streamlit demo application for interactive visualization.

Lucien Martin-Gaffé GitHub Repository: github.com/Lucien-MG/miccai-brats

Residual 3D U-Net with Localization for Brain Tumor Segmentation by Marc Demoustier, Ines Khemir, Quoc Duong Nguyen, Lucien Martin-Gaffé & Nicolas Boutry: link.springer.com/chapter /10.1007/978-3-031-08999-2_33

Inference Performance on the HP ZGX Nano

Segmentation Accuracy of Model Developed on HP ZGX Nano

Results

Methods WT (%) TC (%) ET (%)

Baseline (BCE loss) 90.98 80.24 74.07

Classification Voting Ensemble 1 91.42 80.96 77.07

Regression Voting Ensemble (RVE) 1 91.45 80.98 77.58

BCE-Dice loss (WT & TC) 2 91.34 82.71 77.58

RVE 1 + Localization 4 91.64 82.71 78.26

RVE 1 + ET threshold 100 3 91.45 80.98 78.91

RVE + Localization 4 + ET threshold 400 3 91.64 82.71 78.71

RVE + Localization 4 + ET threshold 600 3 91.64 82.71 80.22

Comparison Speedup (averaged across 10 runs)

NVIDIA GB10 Grace Blackwell Superchip1 vs. NVIDIA GB10 Grace Blackwell Superchip1 CPU Only

16x faster on NVIDIA GB10 Grace Blackwell Superchip1 (0.269s vs. 4.33s)

NVIDIA GB10 Grace Blackwell Superchip1 vs. NVIDIA RTX™ A1000

5.5x faster on NVIDIA GB10 Grace Blackwell Superchip1 (0.269s vs. 1.47s)

Get Started With This Technology

EPITA - School of Engineering & Computer Science 5

I think it deserves special mention that with further CUDA optimizations these speeds on the GB10 GPU could likely be much faster.

https://github.com/Lucien-MG/miccai-brats https://link.springer.com/chapter/10.1007/978-3-031-08999-2_33 https://link.springer.com/chapter/10.1007/978-3-031-08999-2_33

HP provided the customer with an HP ZGX Nano free of charge for evaluation, and their experience is shared here.

1 �Multi-core is designed to improve performance of certain software products. Not all customers or software applications will necessarily benefit from use of this technology. Performance and clock frequency will vary depending on application workload and your hardware and software configurations.

2 �The HP ZGX Toolkit is provided free of charge. Use requires a client device running Windows 11 or Ubuntu 24.04 (or later) with Visual Studio Code installed and host device ZGX Nano. The client device must be x86-based, but otherwise there are no restrictions regarding hardware specifications or device manufacturer. Availability may vary by region and is subject to applicable local laws, regulations, and restrictions.

3 Height excludes feet.

© Copyright 2026 HP Development Company, L.P. The information contained herein is subject to change without notice. HP shall not be liable for technical or editorial errors or omissions contained herein.

c09312070 July 2026.

The HP ZGX Nano G1n AI Station is a compact desktop system designed for AI development and evaluative inference workloads. Built on the NVIDIA GB10 Grace Blackwell Superchip1 with 128GB unified memory, it runs NVIDIA DGX™ OS and, when paired with the HP ZGX Toolkit2, provides a ready-to- use environment for PyTorch, llama.cpp, vLLM, and other AI frameworks. The system delivers powerful AI compute in a 150mm L × 150mm W × 51mm H3 form factor, enabling researchers and developers to iterate on complex models without cloud dependencies or infrastructure complexity.

The HP ZGX Nano is a variant of the NVIDIA DGX Spark™, developed in partnership between HP and NVIDIA.

HP ZGX Nano https://www.hp.com/us-en/workstations/zgx-nano-ai-station.html HP ZGX Toolkit https://www.hp.com/us-en/workstations/zgx-onboard.html NVIDIA DGX Spark™ https://www.nvidia.com/en-us/products/workstations/dgx-spark/

About the HP ZGX Nano AI Station

EPITA - School of Engineering & Computer Science 6

https://www.hp.com/us-en/workstations/zgx-nano-ai-station.html https://www.hp.com/us-en/workstations/zgx-onboard.html https://www.nvidia.com/en-us/products/workstations/dgx-spark/