Fully-automated 1H MRI Thoracic Cavity Segmentation for Hyperpolarized Gas Imaging using a Convolutional Neural Network

Thoracic segmentations are crucial for accurate measurements of normalized lung ventilation, perfusion and gas exchange. Current semi-automated methods are time consuming, require experienced readers, and lack the standardization of fully-automated methods, such as convolutional neural networks. We retrospectively pooled data from 449 healthy and respiratory disease participants, resulting in a 55,000 slice augmented data set to train a dense v-net neural network. The network produced segmentations qualitatively matching semi-automated methods, with high Dice scores and an area under the receiver operating characteristic curve of 0.997. Implementation on the NiftyNet platform permits quick model dissemination for multi-site validation.