gms | German Medical Science

Introduction: Accurate pulmonary function evaluation plays an important role in the diagnosis of respiratory diseases like cystic fibrosis (CF) and chronic obstructive pulmonary disease (COPD). This often requires an individual assessment of each lobe of the lung. Yet, locating lobe fissures on magnetic resonance imaging (MRI) remains a challenging task due to the low density of the images, especially when displaying air-filled structures like lungs.

Computed tomography (CT) is considered the gold standard for diagnosing lung diseases, offering a vast range of algorithms for lung lobe identification. Some of these algorithms achieve Dice scores of up to 0.97 in comparison to experts’ segmentations.

Only a few methods exist for defining lung lobes on MRI, some requiring CT scans for initial lobe segmentation, followed by their transfer to the MRI scan. Tustison et al. [1] proposed a model-based approach where images of lobe segmentations from multiple CT scans are registered onto an MRI scan using a linear affine transform followed by a B-Spline transformation. Pusterla et. al. [2] trained a neural network on CT-based lobe segmentations, showcasing a learning-based approach that leverages advanced ML techniques.

The objective of this work is to develop an automated model-based method for lung lobe identification on MRIs employing a single statistical lung atlas, using advanced image processing techniques.

Methods: The pipeline is implemented with the Insight Toolkit (ITK) in C++. The open-source BodyParts3D dataset [3] contains the used lung atlas, which consists of high-resolution surfaces of each lobe. Initially, the atlas is rotated and rescaled to match the segmentation surface using principal component analysis (PCA). The iterative closest point (ICP) algorithm performs a final alignment of the surfaces.

A nearest neighbor search is used to identify landmarks for a deformable elastic B-spline transformation of the atlas to fit the segmentation. The landmarks are selected from surface points of the segmentation and the atlas lung.

Finally, the registered lobes are converted into binary masks to assign the MRI voxels to their respective lobes.

Results: An initial qualitative analysis of five patient lungs shows promising results. The clinical expert noted slight inaccuracies near lobe boundaries. In particular, the border of the left superior lobe is placed too low, resulting in the supply artery of the left inferior lobe being partially assigned to the superior lobe.

Our radiological partners are currently segmenting lung lobes on MRIs in order to be able to evaluate our approach quantitatively using the SØrensen-Dice-Score and qualitatively through questionnaire.

Discussion: The remaining inaccuracies may stem from the used atlas, as it is based on a single male subject and therefore does not represent the diverse anatomies of human lungs.

Currently, only surface landmarks are utilized in the registration process. Incorporating inner landmarks could also enhance the localization of lobe boundaries. It would be beneficial to examine statistical atlases to determine if their use improves the accuracy of the results. Additionally, the potential use of shape models for that purpose should be explored.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.