Zero-Shot Domain Adaptation in CT Segmentation by Filtered Back Projection Augmentation

Domain shift is one of the most salient challenges in medical computer vision. Due to immense variability scanners’ parameters and imaging protocols, even images obtained from same person scanner could differ significantly. We address computed tomography (CT) caused by different convolution kernels used reconstruction process, critical domain factor CT. The choice a kernel affects pixels’ granularity, image smoothness, noise level. analyze dataset paired CT images, where smooth sharp were reconstructed sinograms with kernels, thus providing identical anatomy but style. Though predictions are desired, we show that consistency, measured as average Dice between on pairs, just 0.54. propose Filtered Back-Projection Augmentation (FPBAug), simple surprisingly efficient approach augment sinogram space emulating kernels. apply proposed method zero-shot adaptation setup consistency boosts 0.54 0.92 outperforming other augmentation approaches. Neither specific preparation source data nor target required, so our publicly released FBPAug (https://github.com/STNLd2/FBPAug) can be plug-and-play module for any CT-based task.