Artificial Intelligence algorithms are not useful if they are not used.

We are surrounded by many examples of researchers trying to achieve the state of the art in developing an algorithm for a particular use-case in radiology.

While there is nothing wrong with doing that, the more we push the limits the greater are the efforts to achieve a marginal increase in model performance.

For one thing, during the development of a lung segmentation model, our lab at Dasa (dasainova) took ~60 hours to train a U-Net to achieve ~0.92 Dice (which is already good). In trying to make the model better, it took ~200 hours more to get ~0.96 Dice, a marginal increase of 0.04.

When do we decide to throw the towel? It depends on the clinical task. It may be the case that a Dice of 0.92 is already good enough for the use-case. If that is true, spending hundreds of hours to get a marginal improvement will only delay the next step, which is the real-world clinical validation of the model (radiologists using it in the context of a clinical trial).

Hence, a good AI algorithm is an algorithm that is used. It may not be the state of the art, but it is good enough and it is available for testing.

When we start thinking about testing a model, one important step is providing the AI result to the radiologist in a timely manner. Sending the AI result back to PACS in an automated way is key.

In that context, Dr. Erickson and I developed Magicians Corner 8: How to connect an Artificial Intelligence to PACS (https://pubs.rsna.org/doi/10.1148/ryai.2021200105…). This editorial article provides the python code that shows how to connect to PACS, query and retrieve a specific study, do the inference of that study in an AI model, then send the results back to PACS.

The paper covers the technical aspect of model integration and can be tested freely in a research environment. If you plan to test it in the production PACS of your institution, be advised there are other issues to be tackled before deploying the model.

For more Radiology AI posts, subscribe to the RSNA AI community: https://communities.rsna.org/communities/community-home?CommunityKey=df6c1c3b-7b69-4dcd-94ce-f1817bd226f9&_ga=2.107715458.1431187753.1605016227-1653207595.1582127521&ssopc=1