Automatic Liver and Lesion Segmentation in CT Using Cascaded Fully Convolutional Neural Networks and 3D Conditional Random Fields

摘要

Automatic segmentation of the liver and its lesion is an important steptowards deriving quantitative biomarkers for accurate clinical diagnosis andcomputer-aided decision support systems. This paper presents a method toautomatically segment liver and lesions in CT abdomen images using cascadedfully convolutional neural networks (CFCNs) and dense 3D conditional randomfields (CRFs). We train and cascade two FCNs for a combined segmentation of theliver and its lesions. In the first step, we train a FCN to segment the liveras ROI input for a second FCN. The second FCN solely segments lesions from thepredicted liver ROIs of step 1. We refine the segmentations of the CFCN using adense 3D CRF that accounts for both spatial coherence and appearance. CFCNmodels were trained in a 2-fold cross-validation on the abdominal CT dataset3DIRCAD comprising 15 hepatic tumor volumes. Our results show that CFCN-basedsemantic liver and lesion segmentation achieves Dice scores over 94% for liverwith computation times below 100s per volume. We experimentally demonstrate therobustness of the proposed method as a decision support system with a highaccuracy and speed for usage in daily clinical routine.