Clara Deploy - Explanation in simple terms


I am experimenting with clara deploy with the help of documentation available. However, I would like to have an ordinary layman explanation of what clara deploy does and how it is useful?

For example, w.r.t Clara Train, I know that it helps radiologists in identifying the region of interest with minimum number of inputs (6 clicks). This is like a semi-supervised/guided approach. This is made possible because Nvidia’s clara platform already has certain pre-trained models. If radiologist is interested to mark Spleen organ, Nvidia clara platform makes it easy because it was trained on spleen dataset. If radiologist wants to mark abnormalities in breast, I understand we have to create our own model.

Please note I have an understanding of the terminologies involved (payload, clara core, DICOM adapter etc) but I am looking for an explanation to help me understand the end-end flow. difference between application development vs workflow development cycle.

Currently we are interested in finding out how clara deploy can help us. We have few imaging datasets (like mammogram etc) and would like to try this with Clara Deploy. So can you guide me on which ones to focus first? As of now, we like to explore the existing features of Nvidia clara train and deploy rather than building our own client (so, we use MITK)

So, can you please explain me in simple terms what does clara deploy do and how we can make effective use of it?

If there are any webinar links or tutorials that would also be helpful

It would really help me kind of figure out things better.

1 Like

Thanks for your interest in Clara deploy and Clara train sdk. Please allow me to explain the sdks from a high level.

Clara deploy sdk:
We provide tools and APIs to make is easy to test existing AI models in the clinic settings. We do this by providing a Dicom listener and writer to interact with Pacs. For this integration hospital IT only need to add a new destination in their dicom router to send images to the m/c that has Clara installed. Clara will then receive the dicom images and trigger the correct work flow(s) then push back the result to Pacs. Clara Deploy also has a render server for visualization and multiple tools to help researcher package the model into the correct format.

Clara Train sdk:
Currently this is still in early access mode. It is split into multiple parts:
Training and transfer learning: Simple APIs that allows researcher to train from scratch or do transfer learning from one of nvidia’s existing models (found on NGC). These API are encapsulating python through configuration files so you don’t even need to know python or ML. For now it is limited to nvidia’s model’s with the next release in a couple of month you would be able to bring your own model.
AI Assisted Annotation: You have a good understanding of this part. Setting it up is straight forward if you follow steps in this blog

Hello Aharouni,

Am I right to understand that Clara Deploy automates the process. Instead of manually collecting the images in CD,doing the processing and analysis, Clara Deploy automates this with the help of Dicom-reader (read images from PACS) to dicom-writer(write output to PACS). But can you please help me understand how does Clara Deploy trigger the workflow based on the image type? I mean which field/info is used by clara deploy to identify the image and the corresponding workflow?


You are correct Clara deploy SDK automates the process through the dicom server, reader and writer. No need to transfer images by CD. All you need to do is configure your pacs router to send images to Clara.

As for “Clara Deploy trigger the workflow based on the image type”
Inside Clara configurations you set up different AEtitles one for each workflow, ex: clara-ct, clara-mri, clara-pet,… and each of them will trigger different workflow.

In order to trigger the workflow according to image type you could do one of 2 things:
1- On your pacs router you configure the logic of checking dicom tag of modality, if modality=CT send to clara-ct, if modality=mri send to clara-mri, etc

2- Have the logic inside your pipeline. So the first stage of your work flow would have this logic then determine which Ai-model to run. This could be a combination of writing your own custom dicom reader and use locks as per nVIDIA’s dicom reader.

The above solutions is for an automatic workflow triggered with no human intervention. Once images go from scanners to dicom router it is automatically sent to your pacs and to Clara for inference. Of course there is always the manual way to have clinicians push a study to once of clara AETitles and wait for the results to come back.

Hope this helps