In the first blog, we discussed the significance of medical imaging, methods of improving the current technologies and processes, challenges being faced during implementation and a proposed solution. In this blog, we will discuss SpringML’s solution in detail and understand its features.
The prominent use cases our solution solves are:
Solution Overview:
- Ready-to-deploy basic end-to-end medical imaging solution.
- Cloud-based medical imaging infrastructure built leveraging Google Healthcare API.
- Smooth interoperability of medical imaging data leveraging PACS.
- Data warehouse to store and manage medical imaging data.
- De-identification functionality.
- Medical image/dataset sharing functionality.
- Embedded Dicom standard image viewer – OHIF.
- Multi-layered security feature with an additional IAP feature.
- Additional Nvidia Clara AIAA capabilities to leverage AI/ML models in medical diagnostics.
Let’s talk about Healthcare API
Cloud Healthcare API bridges the gap between care systems and applications built on Google Cloud. It is a packaged solution to accelerate ingestion, storage, streaming data processing with Cloud Dataflow, scalable analytics with BigQuery, and machine learning with Cloud Machine Learning Engine. Currently supported are HL7v2, DICOM, and FHIR, all of which are widely-used formats for healthcare records and data.
Is it secure though?
Cloud Healthcare API is built using Google’s multilayered security approach that leverages cutting-edge security capabilities, including data-loss prevention tools, precise policy controls, robust identity management, encryption, and many more. Cloud Healthcare API supports HIPAA compliance. Google Cloud is HITRUST CSF certified. Also, Cloud Healthcare API is in scope for Google Cloud Platform’s ISO 27001, ISO 27017, and ISO 27018 certifications. See Google Cloud’s compliance page for more information.
To provide additional security for the customers in terms of data security we use another layer of security with restricting access to the resources. We used Identity Aware Proxy which provides both authentication and authorization which checks if the person accessing is a valid user and has the access to the resource. Moreover, to make data more secure and restricted we use IAM to restrict the bucket access to the specific resource which needs the access.
What other components did we leverage?
Image viewer - OHIF viewer
The Open Health Imaging Foundation (OHIF) Viewer is an open-source, web-based, medical imaging viewer. It can be configured to connect to Image Archives that support DICOMweb and offer support for mapping to proprietary API formats. OHIF maintained extensions add support for viewing, annotating and reporting on DICOM images in 2D (slices) and 3D (volumes). Our solution comes with OHIF’s latest plugin embedded with which we can establish communication to NVIDIA Clara.
NVIDIA Clara
NVIDIA Clara is a healthcare application framework for AI-powered imaging, genomics, and for the development and deployment of smart sensors. Clara Train API is a Unified Python-based API library designed to help accelerate deep learning training and inference. We can both train a new model using the capabilities of AutoML and use a trained annotation or segmentation model to make predictions. Another interesting feature is federated Learning which enables researchers to collaborate and build AI models without sharing private data.
Ok, so now let’s look at how we put these together:
The process can be initiated three ways – manually, automated or scheduled. Data will be loaded/transmitted to Google Cloud Storage from their on-premise systems, and then a dataset will be created with a DICOM store in the Healthcare API. Data will then be imported to the DICOM store from Google Cloud Storage. The de-identification process will be initiated with the tags chosen and stored in a new dataset.
The de-identified data will be exported to Google Cloud Storage for long term cost-effective perseverance. To view and analyze the data before and after analysis, we use the OHIF DICOM viewer hosted on Google App Engine. In addition to removing data for performing predictions on medical conditions, we use the NVIDIA Clara’s solution via OHIF viewer by hosting it in compute engine. Finally, once the analysis is complete, the data from the DICOM stores can be deleted as we preserve the data in Google Cloud Storage.
Medical images require the utmost security and occupy a lot of space. We leverage Google Cloud Storage to store bulk amounts of loads for a longer period and provide the utmost security by using customer-managed or customer-supplied keys to encrypt the data alongside restricting the access of the data. The life of data in the healthcare API is reduced to the amount of time the data requires to be de-identified and analyzed using a DICOM viewer as the cost associated with preserving the data in healthcare API is more compared to Google Cloud Storage.
We provide a granular level control over the PHI data to be de-identified both at the meta-data level and the image level i.e, we can pick only the required amount of info types or profile level categorized by Google. The medical images cannot be directly viewed unlike the other image formats so we used an OHIF viewer to view the image before and after de-identification to both analyze and validate if the required functionality is performed.
Once the PHI information is removed, the next step would be to analyze the data with(out) a 3rd party. For analyzing and predicting if the patient’s scan has any medical problems we have used NVIDIA’s Clara solution for making predictions by loading various models such as covid predictions, tumor predictions, etc from NGC.
In this blog, we have come to understand the important use cases of medical imaging data and how we at SpringML have come up with a solution that creates high value for all stakeholders — healthcare researchers, innovators and providers. This solution is built on Google Healthcare API and leverages the power of NVIDIA Clara to develop and utilize advanced ML models capable of assistive medical diagnosis. For more information on our simple, swift and scalable medical imaging solution, contact us at info@springml.com.
In the third and final blog, we provide a simple cheat sheet of how our solution works and the steps to operate.
