. 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.
- - 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.
Medical computer vision is a way to help doctors look at pictures of our insides to see if we are healthy or sick. Sometimes, the pictures look different because they were taken with different machines or settings. This makes it hard for the computer to understand what it is seeing. The authors found a way to make the pictures look more similar so that the computer can do a better job of understanding them. They tested their method on two sets of pictures and found that it worked really well!
Definitions- Medical computer vision: technology that helps doctors analyze images of our insides
- Domain shift: when images from different sources look different and affect how well the computer can understand them
- Convolution kernels: mathematical tools used in image reconstruction
- Computed tomography (CT): a type of medical imaging that uses X-rays to create detailed pictures of inside the body
- Augmentation: adding extra information to an existing dataset
- Sinogram space: a way of representing CT data before it is reconstructed into an image
- Zero-shot domain adaptation: using a model trained on one set of data to predict on another set without any additional training
- Dice Score: a measure used to evaluate how well two sets of data overlap
- Statistical analysis: using math and statistics to test if results are significant
Medical Computer Vision: Addressing Domain Shift in Computed Tomography Images
Medical computer vision is a rapidly developing field that has the potential to revolutionize healthcare. However, it 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, making it difficult for machine learning models to accurately predict outcomes on unseen data. In this study, the authors address one such source of variability – convolution kernels used in computed tomography (CT) image reconstruction processes – which can affect pixels' granularity, image smoothness, and noise level.
Background
The choice of a convolution kernel affects how CT images are reconstructed from sinograms – collections of projections acquired by rotating an X-ray source around a patient’s body during scanning. Different kernels result in different levels of detail and smoothness in CT images; however, identical predictions are desired regardless of kernel choice. To evaluate consistency between predictions on paired CT images with different kernels, the authors use Dice Score as their 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 significance of results.
Experimental Setup
The authors analyze 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) containing two series each obtained from one physical scanning with different reconstruction algorithms - providing identical anatomy but different styles - resulting in smooth or sharp images depending on kernel choice used during reconstruction process..
Proposed Solution: Filtered Back Projection Augmentation (FBPAug)
To overcome this challenge posed by domain shift due to varying convolution kernels used during image reconstruction process, 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 without requiring specific preparation of source or target domain data . FBPAug is applied within zero shot domain adaptation setup which allows them to boost consistency measured as average Dice score between predictions on pairs from 0.54 up to 0.92 outperforming other augmentation approaches tested alongside it .
Conclusion
In conclusion , FBPAug is a plug-and play module for zero shot domain adaptation tasks that can effectively improve consistency between predictions on paired CT images . The proposed solution was evaluated using two datasets : Mosmed 1110 & Covid private , yielding promising results that outperform other augmentation approaches without requiring any additional preparation steps . This research provides valuable insight into addressing challenges posed by domain shift when working with medical computer vision tasks involving computed tomography imaging .