Medical AI Project Lifecycle

I've been working in the field of machine learning and computer vision for almost 6 years now. But I've been focusing on Medical AI projects for more than 2 years now.

Here's what to expect when working on such projects.

Data

Expect that you will be dealing with a unique type of data when working in the medical field. Types like DICOM, NIFTI or NRRD are amongst the most widely used formats in this field.

This type of data is much more complex than the usual compute vision data that you find in 2D like JPG or PNG.

It involves dealing not just with pixel information, but also with metadata.

Modeling

Expect the modeling to be more complex as well. What I mean by modeling here, is the training and evaluation of your machine learning model.

Since the data is complex to deal with, then the modeling also becomes complex. You will find yourself iterating between the data and the models. This is because, you will usually find yourself choosing between 2 scenarios:

• Either manipulate the data so that you can use classical models found in computer vision (YOLO, Faster RCNN, ...).
• Or you need to use specific models that were developed for such complex data (UNet3D, UNetR, ...).

The answer to the question which scenario should you follow, will depend on your exact use case and production constraints.

Deployment

Deploying machine learning models for medical imaging involves things that you usually do not think about when working with classical 2D images.

You will have to deal with many different constraints, and you'll need to answer many questions such as:

• Should your inference pipeline be synchronous or asynchronous?
• How do you retrieve data to send to your model?
• What should be sent to the server running your model? Raw DICOM files? Zipped DICOMs? NRRD instead of DICOM?
• How do you send the AI models output back to the user?

All of these are things to keep in mind when deploying medical AI models.

Integration

Some people might consider this step as part of the previous step, but I think that it should be its own step. Here's why.

A lot of medical applications that you'll build will only be useful if they are integrated into medical systems like PACS (Picture Archiving and Communication System).

This could be a big task in itself!

PACS expect a standardized DICOM file so any AI model output needs to be converted into such format.

Then there is the problem of how to actually communicate with the PACS. These are challenges that you need to consider if you're developing medical products which will be integrated within hospitals or medical clinics software.

All in all, these are the components of an AI project targeted for medical applications. If you're planning to work in the field, or you're planning to build a company in this space, keep these steps in mind 😉

​

X Post of the Day

Yet another MedTech company getting 510K clearance from FDA!

​

SyntheticMR @Synthetic_MR We are thrilled to announce that our cutting-edge imaging solution, SyMRI 3D, has received FDA 510(k) clearance for clinical use in the United States. 🇺🇸 Read more: https://lnkd.in/dn3FHEGz #SyMRI3D #MedicalImaging #radiology #neuroradiology #mri #mriimaging 6:8 PM • Mar 27, 2024 3 Retweets 14 Likes

​

Share with friends, get cool stuff! Have friends who'd love our newsletter too? Give them your unique referral link (below) and get an awesome reward when they subscribe. [RH_REFLINK GOES HERE] PS: You have referred [RH_TOTREF GOES HERE] people so far ⚡️ by SparkLoop