CT organ segmentation using GPU data augmentation, unsupervised labels and IOU loss

Fully-convolutional neural networks have achieved superior performance in a of image segmentation tasks. However, their training requires laborious annotation of large datasets, as well as acceleration by parallel with high-bandwidth memory, such as GPUs. We show that simple models achieve competitive accuracy for organ segmentation on CT images when with extensive data augmentation, which leverages existing graphics to quickly apply geometric and photometric transformations to 3D image. On 3 mm³ CT volumes, our GPU implementation is 2. 6-8X faster than a-used CPU version, including communication overhead. We also show how to generate training labels using rudimentary morphological, which are efficiently computed by 3D Fourier transforms. We fully-automatic labels for the lungs and bone with semi-automatic ones the liver, kidneys and bladder, to create a dataset of 130 labeled CT. To achieve the best results from data augmentation, our model uses the-over-union (IOU) loss function, a close relative of the Dice loss. discuss its mathematical properties and explain why it outperforms the usual cross-entropy loss for unbalanced segmentation tasks. We conclude that is no unique IOU loss function, as the naive one belongs to a broad of functions with the same essential properties. When combining data with the IOU loss, our model achieves a Dice score of 78-92% for organ. The trained model, code and dataset will be made publicly, to further medical imaging research.