We propose an algorithm to simultaneously prune and train CNNs, leading to networks which have increased generalisability across imaging sites. Through segmentation of data from the ABIDE dataset, we show that through reducing the number of parameters in the network throughout training, we are able to reduce model overfitting, creating a model which is more robust to expect image variations across scanners. We also introduce a novel Targeted Dropout algorithm, which aids the process of model pruning. We demonstrate the approach on a UNet architecture, the basis of nearly all segmentation approaches across medical imaging.
This abstract and the presentation materials are available to members only; a login is required.