Quantifying the importance of spatial anatomical context in cadaveric, non-contrast enhanced organ segmentation

Volumetric segmentation using deep learning is a computationally expensive task, but one with great utility for medical image analysis in radiology. Deep learning uses the process of convolution to calculate voxel level relationships and predict class membership of each voxel (e.g. segmentation). We hypothesize that (1) kidney segmentation in cadaveric, non-contrast enhanced CT images is possible; (2) a volumetric UNet (VNet) architecture will out-perform a 2D UNet architecture in kidney segmentation; and (3) as increasing anatomically relevant information present within the volumes will increase the ability of the system to understand the relationship of anatomical structures, thus enabling more accurate segmentation. In this project we utilized a difficult dataset (cadaveric, non-contrast enhanced CT data) to determine how much anatomical information is necessary to obtain a quantifiable segmentation with the lowest Hausdorff Distance and highest Dice Coefficient values between the output and the ground truth mesh. We used a 70/20/10% training testing and validation split with a total N of 30 specimens. In order to test the anatomical context required to properly segment structures we evaluated and compared the performance of four separate segmentation models: (1) a 2D UNet model that pulled random cross sections from the volumes for training; (2) a 2D UNet model that had the training samples augmented with 3D perturbations for more anatomical context; (3) a 3D VNet with volumetric patching and a padded border to protect against edge artifacts; and (4) a 3D VNet with volumetric patching and image compression by 1/2 the volume with the padded border. Our results show that as anatomical context in the image or volume increases, segmentation performance also improves.