We are happy to present our work on CNN model calibration at the annual German Conference on Medical Image Computing, taking place in Lübeck, Germany. This work was developed in the Bachelor Thesis of Ole H. Martensen, in a collaboration between the DZNE and the Department of Mathmatics at the University of Rostock.
The post-hoc calibration using isotonic regression was already built into our InteractiveVis tool and is currently being revised to provide more stable age-bias correction for dementia probabilities derived from our deep learning model. You can find more details in our paper (to appear soon) and on GitHub https://github.com/CITlabRostock/probcal.
