This is a cheat sheet to guide you through our solution, the features and its respective functionalities.
This UI provides granular control over the Dataset names, DICOM store names, GCS buckets, tags to be identified both at the metadata level and the image level.
Steps to be followed:
- Enter valid Dataset names which would be used to create datasets in healthcare API
- Enter valid datastore names which would be used to create a DICOM store in healthcare API
- The above mentioned naming conventions are really important as these will be used for connecting the on-prem devices as well
- Choose the location which is nearest to you for having a lower latency
- Select appropriate buckets from the GCS for the operations such as importing and exporting the DICOM data
- Choose either the tags or profiles to be de-identified at the image level and profiles to be redacted at image level
- Enter the project id under which the resources are present and the UI is hosted
- Choose the appropriate action based on the requirement i.e,
- Import: imports the data from GCS to the DICOM store created in the Healthcare API
- De-identify: de-identifies the PII and PHI data from the DICOM data available in dataset
- Export: exports the de-identified data from DICOM store to the destination bucket
- To perform any of the above operations use the de-identify button
- To delete the datasets created use the delete dataset option
Dataset a would be created after the pipeline is initiated and once the de-identification process is complete Dataset b will be automatically created with the de-identified dataset.
Steps to be followed:
- If the DICOM store needs to be created explicitly then navigate to Healthcare API section in the GCP billing enabled project
- Create a dataset and then create a store of type DICOM in the dataset created
Dataset creation
DICOM store creation
To view medical images we use OHIF viewer, To the left, the image contains all the PHI data exposed both at the metadata level and at the image level however in the right image all the tags chosen are de-identified.
This UI enables us to load various trained models for predicting various functionalities.
Steps to be followed:
- Navigate to the load balancer service which points to the Clara server running on the compute engine
- Copy the IP address of the appropriate load balancer
- Navigate to the UI and click on Clara present on the right top corner in the header
- Enter the IP address in the following format:
- https://<ext-ip>/docs
- A UI would open under the page which will let us perform all the available options from NVIDIA Clara train
- Load the appropriate model from NGC on which the predictions have to be done via OHIF
The de-identified image can be used by any 3rd party for analysis such as predicting if the scan has a medical problem. Above displayed is a deep grow prediction by NVIDIA Clara v4
Steps to be followed:
- After importing the data from GCS to the DICOM store created in Healthcare API.
- Navigate to the image viewer section in the header to the right top and click it
- Choose the appropriate project-id, location, dataset and store in the OHIF viewer
- Choose the appropriate study on which the dataset of interest resides
- Navigate to the right side on the NVIDIA AIAA plugin
- Enter the IP address of the eternal load balancer under the AIAA server URL
- In the More AIAA settings uncheck the checkbox which points to fetching the data from server (then the image would be sent from OHIF to Clara server)
- Once the model is loaded add a segment and add the appropriate number of data points based on the model i.e,
- Deepgrow model requires points to be pointed continuously like shown in the image above
- Annotation model requires 6 data points
- Segmentation model does not require data points
Related Blogs
