At C-MIMI in Austin this year, one of the talks that caught my attention was on a system called DeepCAT, which is a worklist triage system for mammogram screening. It used deep learning to prioritize mammograms, as well as filtering out low risk images.
It seems that a similar approach would benefit RT image review. Experience with Mosaiq image review users indicates a benefit from anything that can streamline the workflow. A triage model would predict:
This information could be used to:
I‘ve created a simple prototype QA worklist, using the rotation model predictions generated by the SRO decoder ring. It prioritizes a list of synthetic images based on the SRO decoder loss function (in this case, categorical cross entropy).
This same concept could be applied to a clinical image review worklist, but trained with site-specific metric(s). The latent space of patient geometries could be valuable in defining such metrics.