In this work, Gonzalo Mosquera Rojas and his co-authors studied whether uncertainty estimates can help them understand when AI predictions for glioma diagnosis should or should not be trusted. They used a multi-task model that segments the tumor and predicts IDH status, 1p/19q co-deletion and tumor grade from structural MRI, implemented a framework for uncertainty quantification and then looked at how different uncertainty measures behave across these tasks.
They found that uncertainty can help identify less reliable predictions and provide useful information beyond the prediction itself. The results also showed that uncertainty behaves differently depending on the task, highlighting the importance of evaluating it in a task-specific way. More broadly, the framework and evaluation approach are also useful beyond glioma diagnosis and can be applied to other AI applications that use uncertainty quantification.
You can read more here.
