AI’s push into healthcare got a boost yesterday with Nvidia’s release of the Clara Deploy AI toolkit which includes 13 pre-trained models for use in radiology. Clara, you may recall, is Nvidia’s biomedical platform that was introduced at last year’s GTC and described as a “Medical Imaging Supercomputer.” At the time, Nvidia CEO Jensen Huang focused on AI’s capabilities to assist medical imaging tasks. Release of a toolkit with trained models and the capabilities for creating more models is an important forward step.

AI (deep learning and machine learning) will undoubtedly perform many useful tasks in the clinic and bioresearch. Image analysis is seen as one of the early opportunities because of the maturity of AI tools for image processing and because of the tremendous need. Consider the variety of medical images in use as x-ray, MRI, CT, and PET. Fast and accurate interpretation of these images is critical to delivering the right care. AI-assisted annotation has the potential to improve accuracy and great speed up image interpretation.

During his wide-ranging keynote at GTC 2019 yesterday, Huang said, “[Radiologists] used to be able to look at a study for 20 minutes; now they barely have four minutes. The pressure is incredible. And it is the largest operation in the hospital. The question is how do we apply deep learning to enable all of these radiologists and to augment them, so that there’s assistant (AI) sitting next to them helping.”

The new toolkit, developed in collaboration with several prominent biomedical organizations, not only has these 13 pre-trained model, but also features tools for building/training your own models, and sharing them. The already developed models are also available in Nvidia’s container registry. Here’s a snapshot of the Clara SDK’s core capabilities:

• Data Ingestion: Includes a containerized DICOM Adapter interface to communicate with hospital PACS and other imaging systems (both to receive and transmit data)
• Pipeline Manager and Core Services: Provides container based orchestration, resource management & services for TensorRT based inference and Rendered Images Streaming.
• Sample Deployment Workflows: Includes capabilities to define and configure container based workflows using sample workflow with user defined data or modified with user-defined-AI algorithms.
• Visualization Capabilities. Enables the ability to monitor progress and view final results

Huang noted, “There is no way that one institution, one group can possibly train all the neural networks for all of these diseases…We decided instead of being the one company to solve it all, we would help them create tools and put them in the hands of radiologists.”

Huang shared relatively few details about the Clara toolkit but Nvidia posted several blogs (links at end of article) yesterday and are worth reviewing for slightly deeper dives into Clara and the Clara SDK.

Here’s a brief description from one of the blogs (Fast AI Assisted Annotation and Transfer Learning with Clara Train), “Clara Train SDK’s AI Assisted Annotation accelerates the annotation process by enabling application developers integrate the deep learning tools built into the SDK with their existing medical imaging applications, such as MITK, ITK-Snap, 3D Slicer or a custom built application. This is accomplished using a simple API and requires no prior deep learning knowledge. As a result, radiologists can increase their productivity by analyzing more patient data while still using their existing workflows and familiar tools.”

Here an excerpt from another Blog (Clara AI Lets Every Radiologist Teach Their Own AI)

“Transfer learning, another capability in the Clara AI toolkit, adapts existing models to fit local variables. It customizes deep learning algorithms to data that includes local demographics and imaging devices, without having to move or share patient data. As a result, doctors can create models for their own patients with 10x less data than starting from scratch. It takes a significant amount of technical expertise to integrate AI models and applications into hospital IT systems. The toolkit facilitates the integration of AI models into existing radiology workflows using industry standards, like DICOM.”

Here’s a list of relevant blogs with more detail on Clara and Clara SDK:

• Clara AI corporate blog: Clara AI Lets Every Radiologist Teach Their Own AI
• Clara Train dev blog: Fast AI Assisted Annotation and Transfer Learning with Clara Train
• Clara Deploy SDK dev blog: Create, Manage, and Deploy AI-Enhanced Clinical Workflows with Clara Deploy SDK
• MITK release DevNews:  NVIDIA Clara Train Annotation will be integrated into MITK