Cystic fibrosis (CF) is an inherited multi-organ disease, which largely effects the lungs. Repeated bacterial infections and inflammation can result in lung damage, causing most of the morbidity and mortality seen in CF (1, 2). The early detection and monitoring of CF-related lung disease is a prerequisite for optimized care and improved long-term outcomes (3–7).

Recently, chest magnetic resonance imaging (MRI), a radiation-free modality, has shown great promise in assessing structural and functional CF lung abnormalities. Studies have shown chest MRI can detect changes as early as in infancy, and is capable of monitoring disease progression and therapeutic response throughout adulthood (7–14).

To semi-quantitatively assess the severity of lung abnormalities in CF patients, a morpho-functional chest MRI scoring system, also referred to as the Eichinger Score, was developed in 2012 (15). This scoring system includes items for morphological lung abnormalities, as well as perfusion abnormalities (8–11, 15, 16). To automate this scoring process, a critical step is automating the lung segmentation process.

In medical imaging, segmentation refers to identifying an organ or specific tissue of interest by extracting the boundaries and the inner region. This process allows for downstream analysis and extracting important quantitative information within that region. Precise segmentation may support accurate decisions on diagnosis, treatment plans, disease monitoring, and guiding of interventions (17). In the last decade, automated segmentation methods improved in performance and precision, resulting in the possibility of fully automated segmentation in different medical disciplines and imaging modalities (18). Machine learning methods, particularly neural networks, have demonstrated remarkable performance, often outperforming conventional methods, especially when analyzing large datasets (19–21). The nnU-Net, an advanced deep learning framework tailored for medical applications, stands out in its performance (22). It permits the training of networks to perform semantic segmentation with high accuracy and performance, eliminating the need for numerous configuration steps due to its self-configuring training parameters and layer settings. However, difficulties arise when attempting to adapt the nnU-Net to a variety of imaging modalities and diseases. This is particularly challenging when the visual characteristics of pathologic findings deviate significantly from healthy tissue, indicating a change in tissue composition within the same organ (23).

Automated lung segmentation in MR images, especially in the CF population, also have inherent challenges. In MRI, difficulties arise due to the limited spatial resolution and the low contrast between the lungs and the adjacent tissue. In CF patients, breathing artifacts, most notably in young children; cardiac pulsation artifacts; chest growth in children, lung abnormalities displacing air contents, and the deformation associated with disease progression, all contribute to the complexity of the segmentation task (8, 24).

Despite these challenges, many studies are beginning to show promising results incorporating neural networks to automate MRI lung segmentation, even in different underlying pathologies, replacing conventional segmentation approaches (25, 26). Zha et al. applied convolutional neural networks (CNNs) on 3D radial ultra-short echo-times (UTE) oxygen-enhanced MRI in a dataset of 45 subjects (age 10+ with CF, asthma, or healthy) and achieved Dice coefficients of 0.97 and 0.96 for the right and left lung, respectively (27). Furthermore, researchers tested other MRI sequences, such as fast UTE with stack-of-spirals trajectory and matrix pencil decomposition MRI, in CF patients (age 5+) yielding Dice coefficients of 0.96 for children and 0.89 for adults (28, 29).

Notably, Astley et al. tested 2D and 3D nnU-Nets for lung segmentation of patients with varying pulmonary pathologies. In their patient cohort (median age 34 yrs.), analysis of a dataset comprising 809 spoiled-gradient-recalled and UTE MRI scans, even across different vendors, demonstrated a remarkable performance, reaching a median Dice coefficient of 0.96 internally and 0.97 on an external test set (30). To enhance the accuracy of automated lung segmentation, by inclusion of artificially generated images with consolidations, Cristoso et al. reached a Dice coefficient of 0.94 on a cohort of healthy volunteers and patients (31). In a 2023 study of neonates, either healthy or suffering from bronchopulmonary dysplasia, the authors employed CNNs for lung segmentation on quiet-breathing MRI and achieved a Dice coefficient of 0.908 on an internal test set and 0.88 on an independent test set (32). Most recently, a new approach for lung segmentation on healthy adults using thresholding and clustering on an enhanced deep-inspiration-breath-hold reached a Dice coefficient of 0.94 (33, 34). A high benchmark for lung lobe segmentation using pseudo-MRI images derived from CT and three concatenated CNNs achieved a Dice coefficient of 0.95 on a dataset of 100 CF patients over the age of 4.7 years old (35).

To the best of our knowledge, we are the first to demonstrate pediatric lung half segmentations for patients across the entire pediatric age range with different stages of cystic fibrosis using chest MRI on the commonly used sequences BLADE, VIBE, and HASTE. We selected a total of 165 MRI examinations from 84 patients in our internal monocentric CF database. This database contains 1,312 highly standardized annual surveillance MRIs, acquired over more than a decade from 266 patients. Segmentations were created manually by three observers.

Segmentation can play a vital role as a pre-processing step before applying machine learning-based image analysis methods. In our work, lung half segmentation of pediatric MRIs of CF patients using three different sequences, BLADE, VIBE, and HASTE were created utilizing the nnU-Net neural network. A dataset comprising 165 cases, with 55 cases for each of the three sequences, was employed for the training, validation, and testing of the nnU-Net. For each sequence, the nnU-Net was trained individually using a training set of 40 cases and a testing set of 15 cases. For evaluation, the Sørensen-Dice-Score was used in combination with a tailored questionnaire and a slice-based analysis to provide a more detailed insight into the quality of the segmentations.

Overall, the segmentation performance achieved a mean Dice of 0.95 or higher for all sequences and lung halves except for the right lung on the HASTE sequence, which reached a mean Dice of 0.93. With the patient’s age ranging from just a few months to 17 years, the segmentation performance was correlated with age. Generally, it was visible that the segmentation quality stayed constant across all pediatric age classes, further supported by Pearson correlation coefficient r = 0.09. Due to the different disease status of the patients, the global MRI score was correlated with the Dice coefficient. Patients with both lower and higher global MRI score were segmented equally well, which is supported by the Pearson correlation coefficient of r = −0.12. This demonstrates the excellent performance of the nnU-Net for lung lobe segmentation in pediatric chest MRIs within our cohort. The high mean DSC indicates robust segmentation performance, independent of the underlying pathological changes induced by CF in the pediatric stage. An improvement in segmentation performance might be expected with an increased amount of training data (49). However, in segmentation tasks that require precise ground truth annotations, which are extremely time-intensive to generate, necessary trade-offs must be made.

For a qualitative analysis, we provided a questionnaire to the observers for the purpose of evaluating the segmentations manually in addition to the Sørensen-Dice-Score. Consistent segmentation errors in the ventral and dorsal areas of the lung, as well as around the costodiaphragmatic recess were detected. These errors can be caused by the thickness of the image slices, which directly affects the appearance of the tissue. When the slice thickness increases, tissue other than lung tissue becomes included, which may lead to the partial volume effect (PVE) (50). PVE occurs in volumetric imaging, when more than one tissue type is present in a voxel. In cases where the lung parenchyma ends in the middle of the slice, the voxel will have a different shade of gray compared to voxels completely inside or outside the lung. Depending on the patient and amount of non-lung area on the entire slice, this gray level complicates manual and automated segmentation (Figure 7). Furthermore, even the slightest movement in the ventral and dorsal areas may introduce additional artifacts or blurring. While the observers annotated certain areas as lung tissue, the neural network failed to do so. To overcome this challenge, more annotated data could improve the dorsal and ventral segmentation performance. In addition to the questionnaire, the observer conducted a slice-based analysis annotating which slices required manual correction. The analysis revealed that the left lung necessitated more corrections than the right lung.

To achieve a more general evaluation of the trained neural networks, an external dataset from two other centers was segmented and evaluated with the Sørensen-Dice-Score. This was a useful test to explore whether generalization had been achieved, allowing the processing of data from a different source than the training data. Generally, overfitting on the training data is common, leading to very good performances on the training and test set from the same distribution but poor performance on external data. Despite the fact that the Sørensen-Dice-Score for the external dataset did not exceed 0.92, with a range of 0.7–0.92, it suggests that the trained neural network has overall generalizability, given the differences between the two test sets regarding age and number of slices. A comparison of the number of MRI slices in the internal and external test sets reveals notable differences: the internal dataset averages 38 slices per MRI, while the external dataset averages only 19 slices. Since the Dice coefficient is more sensitive to segmentation errors when the overall segmented area is smaller, this difference in slice count and corresponding segmentation area must be taken into account when interpreting the results (51).

Factors such as different MRI scanner specifications, protocols, slice thickness, and resolution affect image quality and therefore segmentation performance. The small number of external MRIs (n = 9) limited the general interpretability. Efforts to improve segmentation performance on the external dataset could include retraining the nnU-Net configuration with external MRI data to reduce segmentation errors. The overall segmentation performances on the three MRI sequences were comparable with existing work in literature. Lung CT scans have been segmented fully automated, reaching high accuracies for lung lobe segmentations with a mean Dice coefficient of up to 0.97 (52). In chest MRI, lung segmentation can be achieved using traditional approaches such as thresholding, but neural network-based segmentation approaches have recently been shown to outperform traditional methods (25). Astley et al. even showed the nnU-Net can be trained to perform well across several sequences, diseases, and vendors reaching a median Dice coefficient of 0.96 on the internal and 0.97 on an independent test set (30). Moreover, their results demonstrated that the 3D-Unet exhibited superior performance compared to the 2D version, which, in turn, outperformed the conventional segmentation approach, spatial fuzzy C-means. In contrast to their work, our study focused on pediatric MRIs of the entire pediatric range of patients with varying degrees of CF disease severity. Efforts toward improving MRI-based lung segmentation include artificially created images to increase robustness in case of severe pathologies (31). For hyperpolarized 129Xe MRI, segmentation performances with a Dice score of 0.929 and above were demonstrated using multiple different methods, highlighting the superiority of the nnU-Net over conventional segmentation methods (53, 54). Neonatal lung segmentations showed a Sørensen-Dice-Score of 0.908 and 0.880 on an independent test set with segmentations automatically by a combination of U-Nets (32).

The resulting segmentations for both the internal and external test sets of the underlying study exhibited variability, yet never attained a Dice coefficient of 1.0. This raises a pivotal question about the criteria for determining whether a segmentation is suitable for subsequent processing or downstream analyses. While a Dice coefficient of 1.0 represents perfect segmentation, striving to improve the coefficient from an already high mean value such as 0.95, may demand a disproportionate amount of time, effort and computational resources. In practice, the pursuit of marginal improvements—bringing the Dice score closer to 1.0—often results in diminishing returns. Such refinements may have minimal impact on the overall effectiveness or accuracy of downstream tasks, particularly when the current segmentation quality is already deemed suitable for clinical decision-making or research purposes. Therefore, it is critical to assess whether the additional time and computational effort invested in further optimizing segmentation is justified, or whether the existing performance is sufficient for the intended applications. When observers segment lungs, they hardly ever reach complete agreement. Segmentation tasks are always dependent on the reader, their experience in the domain, and the tools used. Literature has shown that a Dice coefficient below 0.9 is not uncommon as reader agreement (55). Given that the overarching objective is to automate the Eichinger score, it can be argued that segmentation errors that do not significantly impact the majority of a lung half might be considered acceptable. However, in other research questions, this threshold may have to be set differently, for example, in the context of tumor resection or radiotherapy planning, where segmentations require a higher degree of accuracy (56, 57).

In summary, the results obtained in this study are comparable to those reported in similar studies. To the best of our knowledge, we are the first to demonstrate successful pediatric lung half segmentations for patients with different stages of cystic fibrosis on the MRI sequences BLADE, VIBE, and HASTE.

Our study has some limitations that require discussion. The sample size of 55 cases with corresponding MRIs for each sequence, especially in the age group of patients under 1 year, is relatively small. Compared to the external dataset, all internal cases have larger lungs due to the higher mean age, which could influence the segmentation performance. Moreover, the majority of included patients had a global MRI score of 20 or less. Therefore, it is unclear whether our results are transferable to cohorts of older patients or patients with more advanced lung disease. Future work may focus on this aspect as well as an extension to other sequences.

Recent advancements in this field of research are driving the development of various methods for automated segmentation (19). In the future, it may be valuable to explore these approaches on this dataset and consider expanding the current model to include the remainder of the patients and cases. Especially, since the overall goal of automating the Eichinger score, works toward automated lung lobe segmentations should be explored. Pusterla et al. showed recently that automated lung segmentation with a combination of neural networks is possible with high accuracy (35). Earlier studies demonstrated that segmentation of perfusion maps with a 3D U-Net is an effective approach. However, the evaluation of lung lobes on MRI is challenging due to the difficulty in discerning lobe fissures, if they are visible at all. A lung atlas-based approach, which is independent of age and disease status, may prove advantageous, particularly in light of the findings reported by Tutison et al. regarding the segmentation of the lung (58). With the automated lung segmentation in place, further complex deep learning-based analysis techniques can be applied to assist radiologists in monitoring treatment response, therapy progression, and overall lung health of CF patients, potentially saving time. These results reinforce the efforts toward automated analysis of chest MRIs of patients with cystic fibrosis.

In conclusion, the performance of the nnU-Net in segmenting the lung halves of MRIs from pediatric CF patients demonstrated good agreement with manual segmentations. The segmentation performance of pediatric CF patients does not appear to be significantly influenced by age or disease status.