xAID uses Nebius AI Cloud to train a medical-imaging foundation model on real-world chest and abdominal CT scans. Nebius describes a nearly 580-million-parameter 3D Swin Transformer using inputs up to 256³ resolution. Training on noisy clinical data takes five days per cycle on three NVIDIA H100 GPUs, with uninterrupted compute and MLOps support from Nebius. The dataset includes over one million training scans. An in-house annotation platform helps radiologists correct noisy labels. xAID selected Swin after transformers outperformed alternatives in pre-training experiments. Its shifted attention windows reduce computational load.
The team uses gradient accumulation to fit larger batches within GPU memory limits. Nebius reports GPU utilization of 70–90%. Its healthcare page also describes FP16 training, five days per epoch on noisy clinical data and a goal of detecting more than 70 pathologies. The case study reports a macro-average F1 score of 0.8 overall and 0.86 for five acute conditions relevant to emergency care. Improving accuracy and expanding pathology coverage remain development goals.
In a November 2025 Data in Biotech interview, Nebius’s Ilya Burkov said xAID had supplied models to healthcare providers, with medical imaging teams using them across hospitals and academic environments.
Nebius describes xAID training its CT model on Nebius AI Cloud using three NVIDIA H100 GPUs. The dataset includes over one million training scans, and an in-house annotation platform helps radiologists correct labels. xAID selected Swin after transformers outperformed alternatives in pre-training experiments and uses shifted attention windows to reduce computational load.
Nebius Customer story24 Sept 2026 · Nebius · Web pageNebius describes xAID training its CT model on Nebius AI Cloud using three NVIDIA H100 GPUs. The dataset includes over one million training scans, and an in-house annotation platform helps radiologists correct labels. xAID selected Swin after transformers outperformed alternatives in pre-training experiments and uses shifted attention windows to reduce computational load.
Cloud for Scientific AI and Healthcare
19 Oct 2026 · nebius.com · Web page
Nebius features xAID’s chest and abdominal CT foundation-model training among its healthcare customer examples. It describes 256³ voxel inputs, FP16 precision and accumulated gradients, with a goal of detecting more than 70 pathologies. It specifies five days per epoch on noisy clinical data.
Cloud for Scientific AI and Healthcare19 Oct 2026 · nebius.com · Web pageNebius features xAID’s chest and abdominal CT foundation-model training among its healthcare customer examples. It describes 256³ voxel inputs, FP16 precision and accumulated gradients, with a goal of detecting more than 70 pathologies. It specifies five days per epoch on noisy clinical data.
AI Infrastructure for Biotech — Saturn Cloud and Nebius
5 Nov 2025 · www.corrdyn.com · Web page
Nebius’s Ilya Burkov says xAID used Nebius infrastructure and MLOps support for medical imaging model training on noisy clinical data, with cycles exceeding five days. He says xAID supplied models to healthcare providers and that medical imaging teams use them across hospitals and academic environments.
AI Infrastructure for Biotech — Saturn Cloud and Nebius5 Nov 2025 · www.corrdyn.com · Web pageNebius’s Ilya Burkov says xAID used Nebius infrastructure and MLOps support for medical imaging model training on noisy clinical data, with cycles exceeding five days. He says xAID supplied models to healthcare providers and that medical imaging teams use them across hospitals and academic environments.
xAID provides AI-assisted CT reporting for outpatient imaging centers, teleradiology providers, radiology groups, and small hospitals. Its main service analyzes head, chest, and abdomen CT scans and produces complete draft reports that its in-house European radiologists review before delivery. Customers send imaging studies through DICOM, and reports return directly into their reporting systems through HL7. The customer’s radiologist retains responsibility for the final signature. The service charges per study and supports routine, overflow, and after-hours reporting workflows.
xAID uses Nebius AI Cloud to train a medical-imaging foundation model on real-world chest and abdominal CT scans. Nebius describes a nearly 580-million-parameter 3D Swin Transformer using inputs up to 256³ resolution. Training on noisy clinical data takes five days per cycle on three NVIDIA H100 GPUs, with uninterrupted compute and MLOps support from Nebius. The dataset includes over one million training scans. An in-house annotation platform helps radiologists correct noisy labels. xAID selected Swin after transformers outperformed alternatives in pre-training experiments. Its shifted attention windows reduce computational load.
The team uses gradient accumulation to fit larger batches within GPU memory limits. Nebius reports GPU utilization of 70–90%. Its healthcare page also describes FP16 training, five days per epoch on noisy clinical data and a goal of detecting more than 70 pathologies. The case study reports a macro-average F1 score of 0.8 overall and 0.86 for five acute conditions relevant to emergency care. Improving accuracy and expanding pathology coverage remain development goals.
In a November 2025 Data in Biotech interview, Nebius’s Ilya Burkov said xAID had supplied models to healthcare providers, with medical imaging teams using them across hospitals and academic environments.
Nebius describes xAID training its CT model on Nebius AI Cloud using three NVIDIA H100 GPUs. The dataset includes over one million training scans, and an in-house annotation platform helps radiologists correct labels. xAID selected Swin after transformers outperformed alternatives in pre-training experiments and uses shifted attention windows to reduce computational load.
Nebius Customer story24 Sept 2026 · Nebius · Web pageNebius describes xAID training its CT model on Nebius AI Cloud using three NVIDIA H100 GPUs. The dataset includes over one million training scans, and an in-house annotation platform helps radiologists correct labels. xAID selected Swin after transformers outperformed alternatives in pre-training experiments and uses shifted attention windows to reduce computational load.
Cloud for Scientific AI and Healthcare
19 Oct 2026 · nebius.com · Web page
Nebius features xAID’s chest and abdominal CT foundation-model training among its healthcare customer examples. It describes 256³ voxel inputs, FP16 precision and accumulated gradients, with a goal of detecting more than 70 pathologies. It specifies five days per epoch on noisy clinical data.
Cloud for Scientific AI and Healthcare19 Oct 2026 · nebius.com · Web pageNebius features xAID’s chest and abdominal CT foundation-model training among its healthcare customer examples. It describes 256³ voxel inputs, FP16 precision and accumulated gradients, with a goal of detecting more than 70 pathologies. It specifies five days per epoch on noisy clinical data.
AI Infrastructure for Biotech — Saturn Cloud and Nebius
5 Nov 2025 · www.corrdyn.com · Web page
Nebius’s Ilya Burkov says xAID used Nebius infrastructure and MLOps support for medical imaging model training on noisy clinical data, with cycles exceeding five days. He says xAID supplied models to healthcare providers and that medical imaging teams use them across hospitals and academic environments.
AI Infrastructure for Biotech — Saturn Cloud and Nebius5 Nov 2025 · www.corrdyn.com · Web pageNebius’s Ilya Burkov says xAID used Nebius infrastructure and MLOps support for medical imaging model training on noisy clinical data, with cycles exceeding five days. He says xAID supplied models to healthcare providers and that medical imaging teams use them across hospitals and academic environments.
xAID provides AI-assisted CT reporting for outpatient imaging centers, teleradiology providers, radiology groups, and small hospitals. Its main service analyzes head, chest, and abdomen CT scans and produces complete draft reports that its in-house European radiologists review before delivery. Customers send imaging studies through DICOM, and reports return directly into their reporting systems through HL7. The customer’s radiologist retains responsibility for the final signature. The service charges per study and supports routine, overflow, and after-hours reporting workflows.
