Segmentation is one of the most important tasks in medical image analysis. With the development of deep leaning,fully convolutional networks (FCNs) have become the dominant approach for this task and their extension to3D achieved considerable improvements for automated organ segmentation in volumetric imaging data, such ascomputed tomography (CT). One popular FCN network architecture for 3D volumes is V-Net, originally proposedfor single region segmentation. This network effectively solved the imbalance problem between foreground andbackground voxels by proposing a loss function based on the Dice similarity metric. In this work, we extendthe depth of the original V-Net to obtain better features to model the increased complexity of multi-classsegmentation tasks at higher input/output resolutions using modern large-memory GPUs. Furthermore, wemarkedly improved the training behaviour of V-Net by employing batch normalization layers throughout thenetwork. In this way, we can efficiently improve the stability of the training optimization, achieving faster andmore stable convergence. We show that our architectural changes and refinements dramatically improve thesegmentation performance on a large abdominal CT dataset and obtain close to 90% average Dice score.
展开▼
