Liver lesion segmentation informed by joint liver segmentation

We propose a model for the joint segmentation of the liver and liver lesions in computed tomography (CT) volumes. We build the model from two fully convolutional networks connected in tandem and trained together end-to-end. The first network is trained to produce a representation that is used for liver segmentation. This representation is passed to every layer in the second network, the output of which is used to produce a lesion segmentation. We evaluate the approach on the 2017 ISBI Liver Tumour Segmentation Challenge and place second with a per-volume average Dice score of 0.65.