A 3D Convolutional Neural Network for Volumetric Image Semantic Segmentation

Abstract

This research proposes a novel 3D Convolutional Neural Network (CNN) to perform organ tissue segmentation from volumetric 3D medical images. Accurate and efficient segmentation on the 3D medical image of human organ is a critical step towards disease diagnosis. For volumetric 3D medical image segmentation tasks, the effectiveness of conventional 2D CNNs are reduced due to loss of spatial information. To overcome the obstacles, a 3D CNN that implements the convolution and pooling processes in a 3D space is applied as a substitution to the patch division scheme of 2D CNNs. By using 3D CNN, the image becomes scalable in the spatial direction, allowing accurate image detection with different frame sizes. The 3D CNN implements a cube-by-cube scanning strategy, followed by 3D transformation for each cube in terms of convolving and pooling. The segmentation results of 3D CNN is tested on a 3D OCT image dataset of human thyroid. Experimental results demonstrate that the 3D CNN approach obtains outstanding consistency and accuracy.