Computer Vision News - August 2018

We will start with an overview of the motivation and design goals of NiftyNet, then take a closer look at NiftyNet’s structure, and will conclude with an usage example of CT segmentation. NiftyNet is a TensorFlow-based open-source convolutional neural network platform for research in medical image analysis. NiftyNet’s modular structure is designed for sharing networks and pre-trained models. NiftyNet is a consortium of research organizations (BMEIS, WEISS, CMIC, HIG and UCL), where BMEIS acts as the consortium lead. There is a tendency to consider the medical imaging field simply as a subfield of computer vision generally. However, two aspects make medical imaging unique - - the need for integrating a massive existing field of knowledge (medicine), and the existence of numerous minute details that can make the difference between a really high performance model and one that’s virtually useless. Some examples of the complications of the medical imaging field include: Commonly having to work with 3D data or even 5D data (in the case of MRIs taken at different times), different voxel size and different scales of imaging along different axes, moreover, in sampling the data, you often want to focus on certain areas, as opposed to uniform sampling. The data needs to be referenced per patient, and the difference in datasets (number of images) between patients is of significance, as are image modalities, resolution, etc. 14 We Tried For You: NiftyNet Tool by Assaf Spanier …a TensorFlow-based open-source convolutional neural network platform for research in medical image analysis Computer Vision News