Where to see the results of model prediction when running liver segmentation pipeline?

Hey Guys,

I’m able to run the liver segmentation pipeline and see the output on render server visualization. But, I’m not sure where to check the actual prediction result. where is it stored and how do i view it?

Thanks,

Thanks for the question and the interest in Clara Deploy.

As its overview describes, the liver tumor pipeline produces segmentation of the liver itself as well as the tumor. In the current implementation of the pipeline, the output is sent for the Render Server which provides the use 3D rendering of the segmentation along with original, and additionally, the segmentation image is also wrapped in the DICOM object in a new series, exported back to as configured DICOM receiver.

Please note, Clara Deploy also comes with a DICOM Segmentation Object writer, which generates the proper DICOM Segmentation Object (please see DICOM standards on this IOD and use case) for exporting back the imaging system. This will be incorporated by default in the pipeline in the future releases of the pipeline. Please stay tuned.

Also, Clara Deploy will have the DICOM Structured Report capability as well as the simpler DICOM Encapsulated PDF for classification result reporting. Classification pipelines typically would use these to report back quantitative results (“diagnostic” reports too, though Clara Deploy is not cleared for diagnostics use by any regulatory body)

Regards.

Hey Ming,

Thank you for the response. So if I understand it correctly, currently we can only see the output via render server. The segmentation predictions are not available on any logs generated by Clara.?

The segmentation, wrapped in a new DICOM series of the original study, is sent back to the DICOM receiver, if this is what you are after. The user can then view the segmentation DICOM with his/her preferred DICOM viewer (3D Slicer supports DICOM as well).

In future releases of the Clara Deploy, a DICOM Segmentation object will be sent back to the external DICOM receiver. The DICOM Seg is a single multi-frame DICOM instance, with each frame containing a slice of the labelled segment, with reference to the original DICOM instance.

Hope this helps.

I’m looking for the logs, prediction results’ segmentation mask array (class and pixel details predicted by the model).

Also, are you referring to “dicom-destination-directory” here?
There are files received in this directory but I was unable to open it using 3D Slicer. The files received wasn’t dicom images (it is not “.dcm” file extension)
Error: “virtual bool qSlicerScalarOverlayReader::load(const IOProperties&) failed: missing fileName or modelNodeId property”

I am not sure what happened there. If the external DICOM device is set up (e.g. dcmtk SCP service), and Clara Deploy Dicom Adapter is also set up accordingly to sent DICOM to dcmtk DICOM port and AE Title, the DICOM Seg DICOM instances (dcn files) should show up in the folder on the dmctk side.

I’m just following the quick start guide, at the end it says the “dicom-destination-directory” should have the output dicom images.
https://ngc.nvidia.com/catalog/resources/nvidia:clara:clara_ai_livertumor_pipeline/quickStartGuide.
Do you know if there are there any changes to this documentation?

Btw, I’m also trying to find the actual logs with prediction results’ segmentation mask array (class labels and pixel details predicted by the model). This would’ve been used by the renderservers visualization. Any pointers to this would be helpful.

Please ensure the DICOM Adapter is configured correctly to send back to the dcmtk instance, and there is no firewall rules blocking the traffic.

As I mentioned in a earlier reply, in future releases of the Clara Deploy, a DICOM Segmentation object will be sent back to the external DICOM receiver. The DICOM Seg is a single multi-frame DICOM instance, with each frame containing a slice of the labelled segment, with reference to the original DICOM instance.
Besides, the Clara Management Console will also support downloading the input/output payload of operators, e.g. the segmentation image files from the inference operator.

For now, one can also locate the input/output data in the payloads folder on the server.