The effect of depth context in the segmentation of the colon in MRI volumes

Colonic volume content measurements can provide important information about the digestive tract physiology. Development of automated analyses will accelerate the translation of these measurements into clinical practice. In this paper, we test the effect of data dimension on the success of deep learning approaches to segment colons from MRI data. Deep learning network models were developed which used either 2D slices, complete 3D volumes and 2.5D partial volumes. These represent variations in the trade-off between the size and complexity of a network and its training regime, and the limitation of only being able to use a small section of the data at a time: full 3D networks, for example, have more image context available for decision making but require more powerful hardware to implement. For the datasets utilised here, 3D data was found to outperform 2.5D data, which in turn performed better than 2D datasets. The maximum Dice scores achieved by the networks were 0.898, 0.834 and 0.794 respectively. We also considered the effect of ablating varying amounts of data on the ability of the networks to label images correctly. We achieve dice scores of 0.829, 0.827 and 0.389 for 3D single slices ablation, 3D multi-slice ablation and 2.5D middle slice ablation. In addition, we examined another practical consideration of deep learning, that of how well a network performs on data from another acquisition device. Networks trained on images from a Philips Achieva MRI system yielded Dice scores of up to 0.77 in the 3D case when tested on images captured from a GE Medical Systems HDxt (both 1.5 Tesla) without any retraining. We also considered the effect of single versus multimodal MRI data showing that single modality dice scores can be boosted from 0.825 to 0.898 when adding an extra modality.