Introduction

Phase-Resolved–functional–Lung (PREFUL) MRI enables a free-breathing assessment of pulmonary ventilation and perfusion without contrast agents¹. A crucial factor for the accuracy of this method is the correct segmentation of the lung parenchyma. Currently, the segmentation is automatized by a supervised trained nnU-Net². However, supervised Machine Learning of a segmentation task requires the laborious creation of training data. Furthermore, the resulting model may overfit to the training data and fail for image data e.g. depicting pathologies that were underrepresented in the training dataset³.

A concurrent strategy to detach the dependencies on availability and peculiarities of training data is Zero-Shot-Segmentation (ZSS). ZSS means the segmentation of image content that the segmentation model was not trained on and relies on a foundation model. Recently Meta AI introduced the Segment Anything Model (SAM), a foundation model for image segmentation⁴.

Our study assesses the potential of SAM-based ZSS of lung parenchyma in PREFUL MRI across various acquisition settings in adults and children with pulmonary diseases.

Method

Segmentation from 14 different SAM-configurations were compared with a given ground truth (GT) to 2045 PREFUL-MRI images.

Datasets
The retrospective analysis encompassed image data acquired using different devices and sequence parameters (see Table2). The data was categorized into two groups of children with pulmonary diseases (Cystic fibrosis, non-CF bronchiectasis), two groups of adults with pulmonary diseases (Asthma, COPD) and one cohort of adult healthy volunteers.

Registration
Next, every series of 200-250 images is registered as described by Voskrebenzev et al.⁵ Finally, the mean image of the resulting series is the input of the segmentation step.

Segmentation Ground Truth
GT was defined by manually corrected segmentations from the currently used nnU-Net.

SAM configurations
12 regular point grids, automatically defined using four different values for the configuration parameter Points_per_Side (PPS)⁴ and three different values for Crop_n_Layers (CnL)⁴ were evaluated as segmentation prompts.
Additionally, one respective two seedpoints within the lung represent a further investigated prompt-definition. To prevent interactive definition for every single image, we used the same seedpoints for every image of the dataset. The coordinates of these seedpoints were previously calculated by averaging the centroid coordinates of left and right lung segment from the corresponding GT of the dataset.

Determining the lung segmentation
Since the automatic variant provides multiple segments, only the segment with the highest Dice Similarity Coefficient (DSC) to the GT was considered as the final segmentation result. For the seedpoint-based variants, no distinction was necessary. Figure1 illustrates this process.

Statistical analysis
The best performing ZSS was determined by the median DSC per dataset. ANOVA analysis was used to quantify the DSC differences among the 14 ZSS modes. To assess various MRI setups, we divided each dataset into subgroups defined by combinations of devices and sequence parameters. Finally, the DSC of the SAM approach best performing for the whole dataset (identified using Table1) was calculated for each subgroup in Table2.

Results

As summarized in Table1, SAM provides lung segmentation with a median DSC higher 0.8 in 4/5 groups. The highest median DSC was measured for adult COPD patients and the lowest median DSC for children with cystic fibrosis. In 3/5 groups, seedpoint-based prompts improved the DSC.
ANOVA analysis stated a significant (p<0.001) difference between examined SAM variants, however, there is no decisive impact of vendor or sequence parameter as documented in Table2 and demonstrated in Figure2.

Two kinds of errors could be observed: In some cases, SAM is unable to recognize the lung parenchyma as a separate segment and segments the entire thorax (Figure 3a-b). The second error observed is the insufficient differentiation of the lung from surrounding tissue or vessels (Figure 3c-d).

Discussion

SAM needs no training data to provide a DSC up to 0.87. However, Table1 shows huge differences between SAM configurations for the same data. Furthermore, there is no common best configuration for the examined datasets. This causes an effort to determine the best working approach for a new dataset. A further issue is the recognition of the aimed segment, as SAM may provide not only the segment of interest. Seedpoints can prevent the need for segment recognition but necessitates the definition of suitable coordinates. Nevertheless, it can be assumed that solving these issues is less laborious than manually segmenting sufficient training data.
On the other hand, PREFUL requires a specific segmentation that avoids partial volume effects at the lung boundaries. Such specific requirements are hard to recognize by ZSS. Few-shot segmentation could be a solution and should be evaluated in the future.

Conclusion

Zero-Shot lung segmentation in PREFUL MRI from adults and children with pulmonary diseases is feasible.

Acknowledgements

No acknowledgement found.

References