Cancer Research Advances With NVIDIA AI Enterprise

The Netherlands Cancer Institute (NKI) has been at the forefront of cancer research and treatment since 1913. Composed of an internationally acclaimed research center and a dedicated cancer clinic, NKI puts innovative ideas into action for the benefit of patients.

Recently, NKI started conducting research to see if trained AI models, running on virtualized mainstream servers, could deliver the performance needed to enable more accurate tumor targeting and reduced radiation exposure using cone-beam computed tomography (CBCT). The solution, enabled by NVIDIA’s AI Enterprise software suite, promises to make cancer treatments more effective for more people.

As medical imaging advances, so does the need for increased compute, bandwidth, and storage capabilities. This is especially true in the treatment of cancer, which relies on body imaging with CBCT for accurate planning of targeted radiation therapies.

Unlike higher-quality CT scans, CBCT uses an X-ray tube and a large, flat detector panel that rotates around the patient, capturing data with a cone-shaped X-ray beam instead of the “slices” that CT scanners are typically known for. The data received from these systems is used to reconstruct 3D images for a variety of specialties, including thorax, dental, oral/maxillofacial (mouth, jaw, and neck), and ears, nose, and throat (ENT). CBCT is essential in radiation therapy, where a daily CBCT scan is used for adapting the treatment plan to the current anatomy of the patient, which can change due to weight loss or eating a big meal prior to treatment.

The more accurate the CBCT images, the easier it is for clinicians to localize smaller tumors during radiation therapy and provide more targeted treatments. Accurate CBCT reconstruction would simplify treatment planning, potentially reducing the need for a dedicated planning CT scan.

CBCT images are computed by a reconstruction algorithm using ‌projection data acquired by the scanner. Reconstructing CBCT images, however, is more challenging compared to CT reconstruction, and the image quality given by classical reconstruction methods is poor.

Deep learning reconstruction methods have received a lot of attention from the medical imaging community recently for modalities such as CT and MRI, but the applications to CBCT remain very limited. Training models to reconstruct 3D volumes directly from projection data at clinically relevant resolution and projection count is a memory-hungry task, since a single CBCT volume at 1mm resolution would already occupy 1GB of memory taking gradient information into account.

NKI leveraged a powerful solution developed by NVIDIA, in partnership with VMware, to deploy and manage AI workloads on mainstream 1U/2U servers in a virtualized data center environment, without compromising performance.

NKI researchers were given access to best-in-class AI software—the NVIDIA AI Enterprise suite—certified for the industry’s leading virtualization platform, VMware vSphere. Designed to achieve near-bare-metal performance for AI training and inference, the solution makes larger, more complex AI training and data analytics possible with new scale-out capabilities.

The solution enables high availability for AI workloads, while simplifying infrastructure maintenance, such as consolidation, expansion, and upgrades. VMware vSphere transforms bare-metal servers (including CPU- and GPU-based resources) into centrally managed AI and machine learning infrastructure pools that can quickly provision virtual machines (VMs) and containers that can be accessed from anywhere on demand.

VMware vSphere increases availability, tightens security, streamlines maintenance, and reduces costs to create an agile, efficient, resilient, and secure infrastructure platform that supports existing workloads and next-generation applications.

To evaluate the benefits of updating their infrastructure to the latest GPU architecture, Ampere, the NKI research team developed two different AI models. One model was trained using eight NVIDIA RTX™ 8000 graphics cards, based on the previous generation NVIDIA Turing™ architecture, and graphics cards in a bare-metal environment. The other was trained using two NVIDIA A100 (80GB) graphics cards, based on the newer NVIDIA Ampere GPU architecture, and graphics cards in a virtualized environment with NVIDIA AI Enterprise.

Using eight NVIDIA RTX 8000 graphics cards available on premise, NKI researchers were able to train a low-resolution model that enabled CBCT reconstruction at clinically acceptable 2mm isotropic resolution. This resolution is sufficient for radiotherapy purposes, but it is behind the 1mm resolution given by a dedicated CT scan.

By moving up to two NVIDIA A100 (80GB) graphics cards, the research team was able to fine-tune the low-resolution model and reconstruct full diagnostic images at 1mm resolution. The high-resolution model was initialized from the low-resolution version, and the process of fine-tuning took approximately one month.

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.

The ability to achieve native resolution in near-real time from a CBCT scan makes it possible to provide more accurate treatments to patients, based on their current anatomy when they come in for an appointment.

With a better image, smaller tumors are more likely to be localized reliably on the day of treatment, taking into account displacement of lesions due to weight fluctuations or recent food consumption, and the prescribed radiation dosage can be administered more precisely. In the future, it may be possible to increase treatment accuracy even more by improving tracking of the tumor's motion caused by body movements during respiratory cycles.