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

**Authors:** Talgat Saparov, Anvar Kurmukov, Boris Shirokih, Mikhail Belyaev  
**arXiv:** [2107.08543](https://arxiv.org/abs/2107.08543)  
**Date:** 2021-07-18  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2107.08543v1/

## Key points

- Medical computer vision faces the challenge of domain shift due to variability in scanners' parameters and imaging protocols.
- Different convolution kernels used in the reconstruction process cause variability in computed tomography (CT) images, affecting pixels' granularity, image smoothness, and noise level.
- The authors propose Filtered Back-Projection Augmentation (FBPAug), a simple yet efficient approach to augment CT images in sinogram space emulating reconstruction with different kernels.
- FBPAug is applied in a zero-shot domain adaptation setup and boosts consistency from 0.54 to 0.92 outperforming other augmentation approaches without requiring specific preparation of source or target domain data.
- Two datasets are used for experiments: Mosmed-1110 and Covid-private.
- Dice Score is used as the standard segmentation metric computed as DSC = 2|X∪Y| / |X|+|Y|.
- Statistical analysis using one-sided Wilcoxon signed-rank test and Bonferroni correction is performed to ensure the significance of results.
- FBPAug is a plug-and-play module for zero-shot domain adaptation in any CT-based task that can significantly improve consistency between predictions on paired CT images.

## Summary

. Medical computer vision faces the challenge of domain shift due to the significant variability in scanners' parameters and imaging protocols. Even images obtained from the same person and scanner can differ significantly. In this study, the authors address variability in computed tomography (CT) images caused by different convolution kernels used in the reconstruction process, which is a critical domain shift factor in CT. The choice of a convolution kernel affects pixels' granularity, image smoothness, and noise level. The authors analyze a dataset of paired CT images where smooth and sharp images were reconstructed from the same sinograms with different kernels, providing identical anatomy but different styles. Although identical predictions are desired, they show that consistency measured as the average Dice between predictions on pairs is just 0.54. To overcome this challenge, the authors propose Filtered Back-Projection Augmentation (FBPAug), a simple yet efficient approach to augment CT images in sinogram space emulating reconstruction with different kernels. They apply FBPAug in a zero-shot domain adaptation setup and show that consistency boosts from 0.54 to 0.92 outperforming other augmentation approaches without requiring specific preparation of source or target domain data. The authors use two datasets for their experiments: Mosmed-1110 and Covid-private. Mosmed-1110 contains 1110 CT scans collected from Moscow clinics between March and April 2020, with only 50 scans annotated with binary masks of ground-glass opacity (GGO) and consolidation. Covid-private consists of paired CT studies (189 pairs) of patients with COVID-19 containing two series each obtained from one physical scanning with different reconstruction algorithms. For comparison purposes, they use Dice Score as the standard segmentation metric computed as DSC = 2|X∪Y| / |X|+|Y| for two volumetric binary masks X and Y where |X| is the cardinality of set X. They also perform statistical analysis using one-sided Wilcoxon signed-rank test and Bonferroni correction to ensure the significance of the results. In conclusion, FBPAug is a plug-and-play module for zero-shot domain adaptation in any CT-based task that can significantly improve consistency between predictions on paired CT images.