MENLO PARK, Calif., March 11, 2026 /PRNewswire/ — Subtle Medical, a leader in AI-powered medical imaging software, today announced its participation in NVIDIA GTC 2026, taking place in San Jose, California, March 16-19. At this year’s conference, Subtle Medical will present two technical posters highlighting how SubtleHD™-enhanced data improves downstream accuracy and a foundation model for MRI data standardization.Subtle Medical has been a long-standing partner of NVIDIA, and was the first recipient of the NVIDIA Inception Award in the healthcare category in 2018. Since then, the two companies have collaborated closely to accelerate AI development and deployment in medical imaging, leveraging NVIDIA’s GPU platforms to bring clinically validated AI solutions into real-world healthcare environments.“At NVIDIA GTC 2026, we’re excited to share how advances in accelerated computing and AI infrastructure are enabling faster model development, improved clinical performance, and scalable deployment of medical imaging AI,” said Ajit Shankaranarayanan, PhD, Chief Product Officer at Subtle Medical.Poster Presentations at NVIDIA GTC 20261. A Vision-Language Foundation Model for MRI Data StandardizationSubtle Medical will present a vision-language foundation model designed to standardize MRI metadata across sites, vendors, and protocols, addressing one of the most persistent challenges facing enterprise providers with multiple imaging centers. The team will also evaluate how different training environments, including NVIDIA’s cloud infrastructure (BREV), impact the ability to iterate during model development.Key highlights:
Anna Menyhart-Borroni
Head of Marketing
[email protected]SOURCE Subtle Medical, Inc.
- Trained a vision-language model (SubtleNG™) to harmonize MRI metadata across heterogeneous clinical environments
- Compared training performance across three configurations:
- Local servers with NVIDIA V100 GPUs
- Local servers with NVIDIA A100 GPUs
- NVIDIA cloud infrastructure (BREV) using H100 GPUs
- Local servers with NVIDIA V100 GPUs
- Training on NVIDIA BREV with H100s was 8.8× faster than the best local configuration, enabling full model training in under one day
- Achieved 99% accuracy while reducing metadata heterogeneity by 71%, improving the consistency of cases read by radiologists across different sites.
- Compared dementia classification models trained on standard MR images versus SubtleHD-enhanced images
- Evaluated TensorRT-optimized inference on:
- NVIDIA RTX 2080 Ti
- NVIDIA A100 GPUs
- NVIDIA RTX 2080 Ti
- TensorRT accelerated inference by 3.1× to 5.7×, enabling faster turnaround times for clinical workflows
- SubtleHD-enhanced training data improved model performance, allowing for more accurate classification of dementia cases:
- Standard imaging AUC: 0.951
- SubtleHD-enhanced AUC: 0.968
- Standard imaging AUC: 0.951
Anna Menyhart-Borroni
Head of Marketing
[email protected]SOURCE Subtle Medical, Inc.

Source link













Leave a Reply