Lesson Recap

And this brings our lesson to a close. I hope you learned something valuable and I hope that your newly acquired skills will help you make a greater impact on AI technology for 3D medical imaging!

In this lesson we have covered the following:

• Basics of DICOM networking
• How hospital networks operate and where would AI algorithms fit in
• Requirements for integration of AI algorithms
• Tools for simulating and debugging clinical environments:
  • scripting via DCMTK
  • OHIF - the zero-footprint medical image viewer
  • A deeper-dive into Slicer for annotation
• Medical imaging viewer as radiologist’s tool
• Medical Device Regulations with the US FDA as an example
• Data privacy, HIPAA, and anonymization.

Let me leave you with some useful resources here before we move on to the final project.

Further Resources

• FDA’s guidance for software validation: https://www.fda.gov/media/73141/download
• A very profound analysis done by the University of Cambridge of various regulatory frameworks, as it pertains to algorithms as medical devices: https://www.phgfoundation.org/documents/algorithms-as-medical-devices.pdf
• American College of Radiology has been hosting very good webinars on all things medical imaging and AI.
• An article by American College of Radiology with some greatly structured thoughts on the roadmap for AI development in radiology
• A question that presents somewhat of a challenge to medical device regulators is what to do about machine learning systems that are continuously learning. As you have seen, the traditional regulatory frameworks assume a strict step-by-step process whereby evidence of device performance is collected and submitted once, and if updates are needed, it is submitted again, with 90 days or so review cycles. That doesn’t quite work well for continuously learning systems. FDA has recently published some thoughts on what a framework for such systems could look like.

Vocabulary

• PACS - Picture Archiving and Communication System. An archive for medical images. A PACS product typically also includes “diagnostic workstations” - software for radiologists that is used for viewing and reporting on medical images.
• VNA - Vendor Neutral Archive. A PACS that is not tied to a particular equipment manufacturer. A newer generation of PACS. Often deployed in a cloud environment.
• EHR - Electronic Health Record. A system that stores clinical and administrative information about the patients. EHR system typically interfaces with all other data systems in the hospital and serves as a hub for all patient information. You may also see the acronym “EMR”, which typically refers to the electronic medical records stored by the EHR systems or sometimes used interchangeably with EHR.
• RIS - Radiology Information System. Think of those as “mini-EHRs” for radiology departments. These systems hold patient data, but they are primarily used to schedule patient visits and manage certain administrative tasks like ordering and billing. RIS typically interacts with both PACS and EHR.
• HL7 - Health Level 7. A protocol used to exchange patient data between systems as well as data about physician orders (lab tests, imaging exams)
• FHIR - Fast Healthcare Interoperability Resources. Another protocol for healthcare data exchange. HL7 dates back to the '80s and many design decisions of this protocol start showing their age. You can think of FHIR as the new generation of HL7 built for the open web.
• DIMSE - DICOM Message Service Element. A definition of network message used by medical imaging systems to talk to each other. Often refers to overall subset of DICOM standard that defines how medical images are moved on local networks.
• DICOMWeb - RESTful API for storing and querying DICOM archives. A relatively recent update to the networking portion of the DICOM protocol.
• Application Entity - an actor on a network (e.g. a medical imaging modality or a PACS) that can talk DIMSE messages. Often the abbreviation “AE” is used. An Application Entity is uniquely defined by IP address, Port and an alphanumeric string called “Application Entity Title”.
• SCP - Service Class Provider - an AE that provides services over DIMSE network
• SCU - Service Class User - an AI that requests service from an SCP.
• FDA - Foods and Drugs Administration - a regulatory body in the USA that among other things creates and enforces legislation that defines the operation of medical devices, including AI in medicine. Many regulatory agencies in other countries use regulatory frameworks very similar to that used by the FDA.
• HIPAA - Health Insurance Portability and Accountability Act - key legislation in the USA that among other things defines the concept of Protected Health Information and rules around handling it.
• GDPR - General Data Protection Regulation - European legislation that defines the principles of handling personal data, including health data.