The segmentation of lung volumes constitutes the rst step for most computer{aided systems for lung diseases. CT (Computed Tomography) is the most common imaging technique used by these systems, so fast and accurate methods are needed to for allow early and reliable analysis. In this paper, an e cient and fully automatic method for the segmentation of the lung volumes in CT is presented. This method deals with the initial segmentation of the respiratory system, the posterior extraction of the air tracks, and the nal identi cation of the tow lungs with 3 novel approaches. The system relies only on anatomical assumptions and was evaluated in the context of the VISCERAL Anatomy3 Challenge, achieving one of the best results.
This work presents a novel and fully automatic approach for segmenting the lungs. We rst apply a K{Means [Mac67] clustering of the CT intensities with a xed number of clusters equal to 2 for segmenting the respiratory system. In the second step, the air tracks are removed from the initial segmentation. A novel technique is presented based on the mass{distribution of the lung volumes. The nal step consist of identifying the right and left lung and re ning the nal mask by mathematical morphological operations in 3D. The separation of right and left lungs is challenging when both lungs seem to be connected. In this case, a bidirectional process across the 2D axial slices is applied. It allows to reduce the splitting error due to the information propagated between slices. Once both lungs are identi ed, a re nement in 3D is applied to each lung mask. The entire approach is completely unsupervised and provides an accurate and fast fully automatic segmentation of the lungs. 2
VISCERAL1 Anatomy3 is the benchmark used in the VISCERAL Challenge at ISBI 2015. This benchmark contains a set of medical image series with annotated structures from various modalities. We evaluated our method for segmentation of right and left lung in the modalities of CT and with and withou contrast agent (CTce). A total of 20 training patients in each modality were provided to optimize parameters.
The methods proposed by the participants were executed by the organizers of the challenge in the cloud and tested on a dataset of 10 patients per modality. The test set is not accessible by the participants to avoid possible over tting of the methods. Despite the challenge o ering a training set, the method proposed in this work was set up based on anatomical assumptions and no training was required. Patients from other datasets were used to de ne these assumptions, leaving the training set of this challenge for veri cation purposes. 3
The method presented is composed of three parts: an initial clustering of the CT values for segmenting the complete respiratory system (lungs, trachea and primary bronchi); a process to remove the trachea and primary bronchi; and nally, the identi cation of right and left lung with a re nement of each lung mask (see Fig. 1). Some steps of the process are performed in 2D following the axial dimension of the CT volume, i.e. going through the axial slices.
CT volume
Lungs and air tracks mask
trachea and with no labels iden4fica4on primary and mask bronchi refinement
The proposed method for segmenting the respiratory system is based on the assumption that the latter is the biggest 3D connected air region inside the body. The rst step is to ll the holes in the axial slices by a lling operation [Soi03], where a hole is de ned as an area of dark pixels surrounded by lighter pixels. The resulting image contains a dense{body (see Fig. 2b). Then the absolute di erence between the original and the dense{body image is computed. The resulting 1http://www.visceral.eu/, as of 30 March 2015 image contains values that are clearly larger than 0 in the air regions inside the body, and close to 0 in the other regions (see Fig. 2c). In this new image, a K{Means [Mac67] algorithm with K = 2 is performed, which yields a binary mask (see Fig.2d). Arti cial objects in the CT containing air, such as the plastic bed, may be selected in the clustering, but are removed by analyzing the aspect ratio of the corresponding bounding boxes. Finally, the biggest connected 3D region is used as the initial lung mask. This region showed to include either both lungs connected by the trachea, or only one lung in the case of not being connected by the air track. To deal with this case, the process also selects all air blocks in the same axial slice{range, i.e., in the same slices where the largest 3D region is present and removes the regions that can not be easily connected to the lungs. (a) (b) (c) (d)
Removing trachea and primary bronchi In order to remove the trachea and primary bronchi, the process de nes a plane that divides the 3D image into two parts, leaving an equivalent number of mask{voxels on each side. This process uses the center of mass of the mask obtained in Section 3.1. The plane is used as the reference axis in each slice and the Euclidean distance from every pixel to this axis is computed (see Fig. 3a). Finally, each conencted 2D component (CC) is assigned to the maximum distance found among all its constituting voxels (see Fig. 3b). The regions with a maximum distance to the central axis below a threshold are considered part of the air track and removed. This threshold is dynamically de ned for each slice and patient according to the size of the mask.
(a) (b) (c)
Right{left lung identi cation and mask re nement After removing the trachea and the primary bronchi, two scenarios are present: either the lungs were already 3D{disconnected or they seemed to be merged by the parenchyma, resulting in a single connected 3D component. An algorithm going through the sorted slices was designed to predict the best boundary in those slices where the lungs were connected. First, an initialization of the right (R) and left (L) labels is performed in the rst slice presenting two signi cant CC. Then, the following slices with two CC (so{called 2{CC slices) are consistently labeled by projecting the labels from the previous slice. For the slices presenting only one CC (so{called, 1{CC slices) (see Fig. 4a), the algorithm applies a dilation on the labeled regions from the previous 2{CC slice, and projects them into the region of the current CC. The resulting labeled region contains pixels with one label (R or L), and with two labels (both R and L). This process propagates a boundary assumption to the current slice depending on the previous slice. This propagation results in a di erent labeling if the slices are selected in ascending or in descending order. Hence, the process is executed in both directions and the results are fused. The pixels with double label and the pixels with di erent label due to the double execution de ne a region of con icts, as it is shown in Figure 4b. Then, a K{nearest neighbor algorithm [DHS01] in 3D is applied to decide the best label for each pixel of this region. Other small regions with no label after the procedure are labeled using the adjacent slices. Once both lungs are identi ed, the holes and the cavities are lled for each lung mask independently, achieving the result presented in Figure 4c.
(a) (b) (c) The results shown in this section were provided by the organizers of the VISCERAL Grand Challenge at ISBI 2015. Table 1 shows a subset of the most relevant results. All results are published on the VISCERAL website. The evaluation was performed on the test set detailed in Section 2. The system presented in Section 3 showed to be one of the best algorithms presented in this edition, achieving a minimum Dice coe cient of 0.972 for both lungs in CT and CTce, and a maximum Hausdor distance of 0.052. 5
The method presented in this paper introduces a new method for the extraction of the respiratory system in chest CT volumes. This initial step clearly separates the regions of interest, allowing to apply a fast K-Means clustering with a xed number of 2 clusters. It detects the lung regions in a larger gray{level range than standard thresholding. Moreover, the extraction of the air tracks and the posterior di erentiation of the lungs were done with simple geometric techniques that are computationally inexpensive. The procedures provide a fast system for segmenting the lungs in CT images that can be applied for large datasets. Furthermore, all steps rely on anatomical assumptions [DHS01] [HHR01] 6
This work was partly supported by the Swiss National Science Foundation in the PH4D project (grant agreement 320030{146804).
Shiying Hu, Eric A. Ho man, and Joseph M. M. Reinhardt. Automatic lung segmentation for accurate quantitation of volumetric X{ray CT images. IEEE Transactions on Medical Imaging, 20(6):490{498, 2001.
S.G Armato III and H MacMahon. Automated lung segmentation and computer-aided diagnosis for thoracic fCTg scans. International Congress Series, 1256(0):977{982, 2003. fCARSg 2003. Computer Assisted Radiology and Surgery. Proceedings of the 17th International Congress and Exhibition.
W. Li, S.D. Nie, and J.J. Cheng. A fast automatic method of lung segmentation in ct images using mathematical morphology. In World Congress on Medical Physics and Biomedical Engineering 2006, volume 14 of IFMBE Proceedings, pages 2419{2422. Springer Berlin Heidelberg, 2007.
James MacQueen. Some methods for classi cation and analysis of multivariate observations. In Proceedings of the fth Berkeley Symposium on Mathematical Statistics and Probability, volume 1, pages 281{297. University of California Press, 1967.