05. Requirements for Integration

Heading

Requirements for Integration of AI Algorithms

ND320 C3 L4 06 Requirements For Integration Of AI Algo

Summary

When you start thinking of deploying your AI algorithms, you will want to set some requirements as to data that is sent to these algorithms, and the environment they operate in. When defining those, you may want to think of the following:

  • Series selection. As we’ve seen, modalities typically use C-STORE requests to send entire studies. How are you going to identify images/series that your algorithms will process?
  • Imaging protocols. There are lots of ways images can be acquired - we’ve talked about MR pulse sequences, and there are just physiological parameters, like contrast media or FoV. How do you make sure that your algorithm processes images that are consistent with what it has been trained on?
  • Workflow disruptions. If the algorithm introduces something new into the radiologists' workflow - how is this interaction going to happen?
  • Interfaces with existing systems. If your algorithm produces an output - where does it go? What should the systems processing your algorithm’s output be capable of doing?

Series Selection

An AI solution can quantify the severity of knee ligament damage on a Proton Density (PD) MRI volume. You deploy it between an MRI scanner and PACS to generate PDF reports out of any suitable series, sending them by email. The MRI scanner is used by the radiology department for different scans - abdominal, brain, etc. What strategies would help identify the series to run your algorithm on?

SOLUTION:
  • Looking at [Series Description](http://dicom.nema.org/medical/dicom/2020a/output/chtml/part03/sect_C.7.3.html) tag
  • Checking [Body Part Examined](http://dicom.nema.org/medical/dicom/2020a/output/chtml/part03/sect_C.7.3.html) tag
  • Building a classifier that will look at MR series and predict if this is a PD image of a knee