Real-time Analytics
The Netherlands Cancer Institute needed faster, higher-resolution cone-beam CT reconstruction so clinicians could adapt radiotherapy to a patient's current anatomy, but memory-heavy 3D reconstruction models and CPU-era infrastructure made clinical turnaround too slow.
Reported impact
NKI reduced 1 mm CBCT reconstruction time from 14 minutes on a bare-metal RTX 8000 system to five minutes on a single A100, rendered 2 mm images in about 60 seconds, and created a more maintainable shared GPU environment for future radiotherapy AI workloads.
NKI evaluated whether modern AI reconstruction could run on virtualized infrastructure rather than isolated bare-metal systems by combining NVIDIA AI Enterprise, VMware vSphere, and NVIDIA A100 80GB GPUs. The research team developed CBCT reconstruction models in both older RTX 8000-based environments and newer Ampere-based virtual machines, then used the larger A100 memory footprint to fine-tune high-resolution 1 mm models that previously required downsampling. vSphere pooled CPU and GPU resources into centrally managed infrastructure, while AI Enterprise provided the software foundation for training and serving the models. That stack let NKI reconstruct 2 mm diagnostic images in as little as 60 seconds and 1 mm images in about five minutes, bringing clinically relevant, AI-driven image reconstruction closer to routine adaptive radiotherapy planning.
Projection data from CBCT scans is fed into AI reconstruction models running in vSphere-managed virtual machines. NVIDIA AI Enterprise and A100 GPUs process the volumes at diagnostic resolution, and the reconstructed 3D images are returned to clinicians for more accurate tumor targeting and treatment planning.
Diagnostic images could be rendered in as little as 60 seconds at 2mm resolution and in as little as five minutes at 1mm resolution.
NVIDIA AI Enterprise software and NVIDIA A100 GPUs were used in a VMware vSphere environment.
The inference speed at 1mm resolution was also greatly improved by moving to the A100: A full volume at 1mm resolution took just five minutes to reconstruct on a single GPU, compared to 14 minutes on the bare-metal RTX 8000 machine.
With AI models and hardware from two or three years ago, we'd have to down-step the resolution to free up memory for training.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
The 2024 key figures page lists Medisch specialisten 284, Verpleegkundigen (incl. zorgassistenten) 405, Medewerkers onderzoek 747, and Overige medewerkers 2029; together these categories total 3465 staff.
Related
NVIDIA Sample Apr 23, 2026 Biofy Hospitals needed pathogen identification and antibiotic-resistance analysis much faster than five-day culture and antibiogram workflows, but CPU-bound genomic processing and limited AI expertise made it hard to turn… 100,000+ synthetic DNA sequences generated by the biology model Abby · H100 Tensor Core GPUs · NVIDIA CUDA-accelerated libraries NVIDIA Sample Archived · Feb 9, 2026 Clinomic Clinomic needed to consolidate ICU patient data, reduce documentation burden, and run bedside analytics and telemedicine functions without relying on cloud connectivity that could introduce latency or operational risk… 143% increase in patient-care capacity NVIDIA Jetson AGX Xavier · NVIDIA TensorRT NVIDIA Sample Feb 2, 2026 UNSW Sydney UNSW Sydney and the NSW Telestroke Service needed fast, quantitative stroke imaging for regional hospitals, but manual hemorrhage and infarct segmentation took too long and made it hard to guide treatment decisions in… 20 minutes to seconds Auto3DSeg · MONAI Label · NVIDIA GPUs on AWS 