V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image\n Segmentation

Abstract

Convolutional Neural Networks (CNNs) have been recently employed to solve\nproblems from both the computer vision and medical image analysis fields.\nDespite their popularity, most approaches are only able to process 2D images\nwhile most medical data used in clinical practice consists of 3D volumes. In\nthis work we propose an approach to 3D image segmentation based on a\nvolumetric, fully convolutional, neural network. Our CNN is trained end-to-end\non MRI volumes depicting prostate, and learns to predict segmentation for the\nwhole volume at once. We introduce a novel objective function, that we optimise\nduring training, based on Dice coefficient. In this way we can deal with\nsituations where there is a strong imbalance between the number of foreground\nand background voxels. To cope with the limited number of annotated volumes\navailable for training, we augment the data applying random non-linear\ntransformations and histogram matching. We show in our experimental evaluation\nthat our approach achieves good performances on challenging test data while\nrequiring only a fraction of the processing time needed by other previous\nmethods.\n

Ähnliche Arbeiten