Modern radiotherapy techniques are vulnerable to delineation inaccuracies owing to the steep dose gradient around the target. In this aspect, accurate contouring comprises an indispensable part of optimal radiation treatment planning (RTP). We suggest a fully automated method to segment the lungs, trachea/main bronchi, and spinal canal accurately from computed tomography (CT) scans of patients with lung cancer to use for RTP. For this purpose, we developed a new algorithm for inclusion of excluded pathological areas into the segmented lungs and a modified version of the fuzzy segmentation by morphological reconstruction for spinal canal segmentation and implemented some image processing algorithms along with them. To assess the accuracy, we performed two comparisons between the automatically obtained results and the results obtained manually by an expert. The average volume overlap ratio values range between 94.30 ± 3.93% and 99.11 ± 0.26% on the two different datasets. We obtained the average symmetric surface distance values between the ranges of 0.28 ± 0.21 mm and 0.89 ± 0.32 mm by using the same datasets. Our method provides favorable results in the segmentation of CT scans of patients with lung cancer and can avoid heavy computational load and might offer expedited segmentation that can be used in RTP.
Computed tomography (CT) scans are primarily used for diagnostic purposes; however, they may also be used in radiation treatment planning (RTP). Detailed CT scans of patients with cancer are acquired for RTP purposes and used for localizing the tumor and organs at risk (OARs). Optimal RTP requires precise definition of target and critical structures to achieve the best radiotherapeutic outcomes in terms of toxicity and cure. Modern radiotherapy techniques such as intensity modulated radiation therapy (IMRT) are vulnerable to delineation inaccuracies due to the steep dose gradient around the target. In this aspect, accurate contouring is extremely important and comprises an indispensable part of RTP in the modern era.
The delineation procedure is traditionally performed by an expert in radiation oncology with meticulous assessment of CT images of a given patient followed by manual construction of the structure set by drawing two-dimensional contours of every structure in consecutive CT slices. Depending on the tumor site and number of slices to be contoured, delineation of the target and OARs for precise RTP can be quite time-consuming and labor-intensive. Moreover, there is generally no consensus on accurate contouring of target and OARs to guide this critical component of RTP because of the fuzziness of image objects. Contouring procedures performed by different experts may differ substantially and a single expert may even contour the same CT image differently on consequent occasions, referred to as “interobserver variability” and “intraobserver variability,” respectively [
Automated segmentation algorithms are increasingly used in RTP to optimize the delineation procedure. Robust algorithms significantly expedite the contouring and improve consistency and concordance. The implementation of these sophisticated methods in radiation oncology practice may have implications particularly in busy clinics.
Segmentation of the lungs, trachea/main bronchi, and spinal canal plays a central role in RTP for lung cancer. Most of the lung segmentation approaches [
de Nunzio et al. [
Yim and Hong [
Wang et al. [
Sluimer et al. [
Prasad et al. [
Shape models [
Airway tree segmentation is critical in correcting the results of lung segmentation, that is, elimination of the external airways from the segmented lungs. Most of the airway tree segmentation methods use region growing and morphological operators applied on the density values [
Methods for segmenting the spinal cord and the spinal canal include knowledge-based approach [
Banik et al. [
Haas et al. [
In this study, we suggest a fully automated method to segment the lungs, trachea/main bronchi, and spinal canal accurately from CT scans of patients with lung cancer to be used for RTP purposes. Our method consists of three processes. First, the body region of the patient in a CT image is segmented by elimination of the background. Second, rough segmentation of the lung fields, segmentation and elimination of the trachea/main bronchi, the lung fields correction, the right and left lung seperation steps, and a postprocessing step for inclusion of excluded pathological areas into the segmented lungs are realized, respectively. Third, the vertebra and finally the spinal canal are segmented by means of the fuzzy segmentation algorithm. Within these processes, a new algorithm for inclusion of excluded pathological areas into the segmented lungs, a modified version of the fuzzy segmentation by morphological reconstruction for spinal canal segmentation, and the well-known image processing algorithms were used.
Section
CT scans of 10 patients undergoing radiotherapy at the Department of Radiation Oncology, Gülhane Military Medical Academy for primary lung cancer, were used in our study. Informed consents of all patients were taken before CT acquisition at the dedicated CT-simulator (GE Lightspeed RT, GE Healthcare, Chalfont St. Giles, UK). Slices are 512 × 512 pixel, 16-bit gray level matrices; and pixel size ranges between 0.76 mm and 1.27 mm. The average number of slices per scan is 100 (range: 79 to 129 slices) while slice thickness ranges between 2.5 mm and 5 mm.
In addition to the 10 CT scans of the 10 patients with lung cancer, we also used 10 different thoracic CT scans from the Lung Image Database Consortium (LIDC) [
To compare our method with other methods, we implemented the algorithms proposed in other studies [
Figure
Workflow of the study.
The body region is segmented from a CT image by thresholding. First, the Hounsfield Unit-HU (range: −175 to 750) [
Segmentation of the body region: (a) original CT slice, (b) segmented body region in white, and (c) body contour in red superimposed on the original slice.
As shown in Figure
Segmentation of the lungs: (a, b) original CT slices, (c, d) rough segmentation of the lung fields of (a, b) in white, (e) lungs in white after eliminating the bronchi from (c), (f) lungs in white after removing intestine from (d), and (g, h) lung contours in red superimposed on the original slices.
The body segments of each slice are thresholded using the value of −300 HU [
A three-dimensional region growing is used to segment the trachea and main bronchi. Within the body segments of the first (upper) few slices, all the pixels with density values lower than −900 HU are labelled as air-filled since air has very low HU values around −1000 in CT slices. By using the connected component analysis, the air-filled region closest to the center of the corresponding body segment with the maximum area is labelled as the trachea.
Using the center pixel of the trachea region as a seed, three-dimensional region growing is applied repeatedly with increasing values of the threshold. Here, initial value of the threshold is −900 HU and we take 64 HU as a value of increment. In order to find the center pixel of the trachea, average location of the pixels in the trachea is calculated.
If the segmented structures have a total volume at least twice the structures segmented with the previous threshold, it is considered that the growing region penetrates through the bronchial wall and enters into the lung parenchyma. In this case, value of the increment is reduced by half. This operation is terminated when the increment reaches the value of 1 HU and leakage into the lung field is detected synchronously. Figure
Results of three-dimensional region growing for a CT slice: (a) original CT slice, (b) used threshold which is −836 HU, (c) used threshold which is −772 HU, and (d) used threshold which is −708 HU.
To include the airway wall with higher density values than the air-filled region (lumen), we apply morphological dilation with a 3 × 3 disk-shaped structuring element [
Segmentation of the trachea/main bronchi: (a) original CT slice, (b) segmented trachea/main bronchi in white, and (c) trachea/main bronchi contour in red superimposed on the original slice.
We fill the holes stemming from vessels, nodules, tumors, or other high density pathologies that are inside the lung fields using a hole-filling algorithm. Finally, morphological closing with a 3 × 3 disk-shaped structuring element [
A three-dimensional evaluation of the CT scan is performed to remove intestine that has similar density values as the lungs. Intestine appears in the lower (caudal) slices of the scan. All the rough lung regions smaller than 200 mm2 are eliminated and connectivity is checked within the remaining lung regions by means of three-dimensional region growing [
If a slice contains a lung region wider than half of the width of the body region, then separation of the connected lungs is performed by identifying the anterior and posterior junctions as follows.
Find the bounding box (BB) of the lung region and determine the boundaries of BB as
In order to generate the anterior and posterior junction lines, determine a region of interest (ROI) using the border definitions below:
Find the greatest nonlung component that is in the middle upper part of the ROI.
For anterior junction line, find the pixel of the nonlung component obtained from the previous step with minimum row position, that is, nearest pixel to the
Compare the density values of
If
For posterior junction line, find the pixel of the nonlung component obtained from Step
Compare the density values of
If
The result is shown in Figure
The right and left lung separation: (a) original CT slice, (b) connected lungs in white, and (c) right lung contour in yellow and left lung contour in red superimposed on the original slice.
Although the lung segmentation based on thresholding is simple and quite fast, it may fail in case of lungs with large tumors of high density since a significant contrast between the lungs and the surrounding tissues does not exist. In such circumstances, morphological operations like closing may not be sufficient to correct the borders of the lungs.
We propose a three-stage approach to include pathological areas, that is, tumors that are in relation with the borders of the lungs and are excluded by segmentation in the previous steps, into the lungs. This approach is based on obtaining the intersection of interpolated and propagated lungs.
Apply an interpolation procedure to the lungs obtained from the previous subprocess.
Create an empty mask, that is, a two-dimensional matrix the same size as the slices.
Starting from the first (upper) slice to the last (lower) slice of CT scan, if a pixel
For each pixel
Label all the pixels
Propagate the right and left lungs obtained from the previous subprocess separately.
Find the border pixels of the lung.
For each border pixel, find the nonlung pixels within the neighborhood (7 × 7) centered at the current border pixel and label them as candidate pixels.
For each candidate pixel, if more than half of the pixels within the neighborhood (7 × 7 × 7) centered at the current candidate pixel are labelled as lung, schedule the current candidate pixel for inclusion into the lung.
If there is any scheduled candidate pixel, label them as lung and go to Step
Get the intersection of the lungs that resulted from Stages
Herein, neighborhood of 7 × 7 × 7 region was chosen experimentally by testing this approach on the CT scans from the Department of Radiation Oncology, Gülhane Military Medical Academy. These CT scans belong to 10 patients with limited-stage small cell lung cancer. Each of the patients has one tumor. Gross tumor volumes (GTVs) are 36.7, 16.4, 85.8, 33.0, 28.7, 40.8, 66.5, 21.7, 45.5, and 93.2 cc. Figure
Effects of different neighborhoods used in Section
Inclusion of excluded pathological areas: pathological areas contours in red superimposed on the original slice.
In order to see results of this approach in case of juxtapleural nodules, we used 10 CT scans from the LIDC. The CT scans from the LIDC do not include large tumors but comprised a total of 12 juxtapleural nodules that are 5.0–10.0 mm in diameter.
As can be seen from Section
Fuzzy segmentation approach is examined in segmentation of the spinal canal. This process includes two subprocesses: segmentation of the vertebra and fuzzy segmentation of the spinal canal.
Since the spinal canal is the space in vertebra through which the spinal cord passes, vertebra must be segmented initially. Bones have higher density values than other structures so that the body region is thresholded with a value of 145 HU [
Find the bounding box (BB) of the body segment and determine the boundaries of BB as
Determine the coordinates of the center point (CP) of BB as
Determine a region of interest (ROI) for vertebra detection using the border definitions below:
Label the bone segment that overlaps with the ROI as vertebra and create a binarized vertebra image for the corresponding slice.
In the vertebra image detect the pixels that have nonzero gradient magnitude values and label them as vertebra in order to include the missing parts.
Banik et al. [
We use a modified version of the fuzzy segmentation by morphological reconstruction presented in aforementioned study [
The steps of the fuzzy segmentation approach that we propose are as follows.
Determine if the slice has a spinal canal enclosed completely by the vertebra or not. To do this, find the bounding box of the vertebra. In the bounding box of the vertebra, find the pixels that are not labelled as bone. Using connected component labelling, detect the segments that the pixels form. If the slice has a segment that is not connected to the boundaries of bounding box of the vertebra, it is understood that the slice has a spinal canal enclosed completely by the vertebra.
If so, label the nonbone region enclosed by the vertebra as the spinal canal and then go to Step
If not, take CP of the spinal canal in the previous (following) slice as a seed.
Take a window of size 11 × 11 pixels centered at the coordinates of CP in the current slice. Detect the pixels in the defined window having density values in the range of
Reconstruct the fuzzy region according to the fuzzy membership function, namely, the unnormalized Gaussian function, using
Binarize the fuzzy region using 0.5 as the threshold value.
Using the connected component labelling, find the segment with the maximum area in the thresholded fuzzy region and label it as the spinal canal.
Morphologically open [
Take the following (previous) slice and go to Step
The result of this process is shown in Figure
Segmentation of the spinal canal: (a) original CT slice in which the vertebra encloses the spinal canal completely, (b) original CT slice in which the vertebra does not enclose the spinal canal completely, (c, d) segmented spinal canals of (a, b) in white, and (e, f) spinal canal contours in red superimposed on the original slices.
The proposed segmentation method was implemented in Matlab R14 and tested on a PC with 1.73 GHz processor and 1.0 GB RAM. Also, we implemented the methods proposed in other studies [
Automatic and manual segmentation of the lungs, trachea/main bronchi, and spinal canal: contours in yellow show automatic segmentation results and contours in green show manual segmentation results.
The contours of automatically segmented structures: body contours in yellow, left lung contours in blue, right lung contours in red, trachea/main bronchi contours in green, bone contours in magenta, and spinal canal contours in cyan superimposed on the original slices.
Similar to other studies [
In order to solve the undersegmentation problem caused by pleural nodules and pulmonary vessels contacting the lung boundary, de Nunzio et al. [
In this study, we propose an original method to obtain pathological areas in the segmented lungs as described in Section
Banik et al. [
To assess the accuracy, we performed two comparisons between the automatically obtained results and the gold standard. Here, the results obtained manually by an expert were used as gold standard. An expert radiation oncologist manually delineated the body region, right and left lung, trachea/main bronchi, and spinal canal in consecutive slices of all CT scans by mouse dragging at a dedicated contouring workstation using Advantage SimMD simulation and localization software (Advantage SimMD, GE, UK) [
The first comparison was performed by computing the volume overlap ratio (VOR) [
Volume of a structure is computed by taking the product of total number of pixels labelled as that structure, pixel dimensions (width and height), and slice thickness.
As the second comparison, surface distance evaluation [
For a given structure, the average symmetric surface distance (ASD), RMS symmetric surface distance (RMSD), and maximum symmetric surface distance (MSD) are calculated using the following equations:
Comparison of the results of our method and the other methods with the gold standard is shown in Tables
Comparison of VOR (%) for the 10 CT scans from the LIDC.
OARs | Method | ||||
---|---|---|---|---|---|
Our method |
de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 99.14 ± 0.33 | 99.12 ± 0.38 | — | — | — |
Right lung | 99.07 ± 0.30 | 99.03 ± 0.42 | — | — | — |
Lungs (together) | 99.11 ± 0.26 | — | 99.15 ± 0.30 | 98.90 ± 0.30 | — |
Trachea/main bronchi | 96.91 ± 1.47 | 97.06 ± 1.30 | 97.68 ± 1.42 | 95.61 ± 1.65 | — |
Spinal canal | 97.19 ± 2.72 | — | — | — | 97.11 ± 2.35 |
Comparison of the average symmetric surface distance (mm) for the 10 CT scans from the LIDC.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 0.31 ± 0.10 | 0.44 ± 0.19 | — | — | — |
Right lung | 0.27 ± 0.17 | 0.34 ± 0.12 | — | — | — |
Lungs (together) | 0.29 ± 0.03 | — | 0.32 ± 0.11 | 0.34 ± 0.20 | — |
Trachea/main bronchi | 0.34 ± 0.14 | 0.36 ± 0.05 | 0.39 ± 0.13 | 0.59 ± 0.15 | — |
Spinal canal | 0.28 ± 0.21 | — | — | — | 0.35 ± 0.09 |
Comparison of the RMS symmetric surface distance (mm) for the 10 CT scans from the LIDC.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 0.61 ± 0.22 | 0.79 ± 0.23 | — | — | — |
Right lung | 0.63 ± 0.38 | 0.82 ± 0.51 | — | — | — |
Lungs (together) | 0.67 ± 0.18 | — | 0.71 ± 0.22 | 0.89 ± 0.37 | — |
Trachea/main bronchi | 0.66 ± 0.35 | 0.69 ± 0.20 | 0.81 ± 0.24 | 1.13 ± 0.41 | — |
Spinal canal | 0.60 ± 0.43 | — | — | — | 0.73 ± 0.38 |
Comparison of the maximum symmetric surface distance (mm) for the 10 CT scans from the LIDC.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 1.76 ± 0.66 | 1.83 ± 0.99 | — | — | — |
Right lung | 1.93 ± 1.03 | 2.02 ± 0.87 | — | — | — |
Lungs (together) | 2.08 ± 1.15 | — | 2.23 ± 1.36 | 2.76 ± 1.90 | — |
Trachea/main bronchi | 2.55 ± 1.44 | 2.88 ± 0.97 | 3.03 ± 1.65 | 3.51 ± 2.24 | — |
Spinal canal | 2.67 ± 1.89 | — | — | — | 3.78 ± 2.00 |
Comparison of VOR (%) for the 10 CT scans of the 10 cancer patients.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 98.70 ± 1.32 | 96.50 ± 0.91 | — | — | — |
Right lung | 98.70 ± 0.86 | 96.30 ± 1.12 | — | — | — |
Lungs (together) | 98.70 ± 1.27 | — | 97.10 ± 1.02 | 95.40 ± 1.82 | — |
Trachea/main bronchi | 94.30 ± 3.93 | 94.60 ± 3.35 | 94.60 ± 2.87 | 93.00 ± 3.63 | — |
Spinal canal | 96.50 ± 3.67 | — | — | — | 96.70 ± 3.59 |
Comparison of the average symmetric surface distance (mm) for the 10 CT scans of the 10 cancer patients.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 0.73 ± 0.36 | 0.90 ± 0.51 | — | — | — |
Right lung | 0.77 ± 0.48 | 0.93 ± 0.71 | — | — | — |
Lungs (together) | 0.63 ± 0.32 | — | 0.94 ± 0.57 | 0.99 ± 0.73 | — |
Trachea/main bronchi | 0.89 ± 0.72 | 0.50 ± 0.23 | 0.55 ± 0.31 | 0.71 ± 0.22 | — |
Spinal canal | 0.57 ± 0.41 | — | — | — | 0.52 ± 0.39 |
Comparison of the RMS symmetric surface distance (mm) for the 10 CT scans of the 10 cancer patients.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 1.33 ± 0.48 | 1.77 ± 0.63 | — | — | — |
Right lung | 1.48 ± 0.75 | 1.51 ± 0.89 | — | — | — |
Lungs (together) | 1.30 ± 0.93 | — | 1.92 ± 1.13 | 2.16 ± 1.24 | — |
Trachea/main bronchi | 1.52 ± 0.86 | 1.23 ± 0.55 | 1.38 ± 0.77 | 1.56 ± 0.97 | — |
Spinal canal | 1.13 ± 0.78 | — | — | — | 1.02 ± 0.62 |
Comparison of the maximum symmetric surface distance (mm) for the 10 CT scans of the 10 cancer patients.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Left lung | 8.42 ± 3.48 | 9.16 ± 3.01 | — | — | — |
Right lung | 8.23 ± 4.12 | 10.02 ± 4.65 | — | — | — |
Lungs (together) | 8.57 ± 2.88 | — | 10.67 ± 2.78 | 11.13 ± 4.34 | — |
Trachea/main bronchi | 11.78 ± 4.35 | 11.21 ± 3.98 | 11.28 ± 4.74 | 12.96 ± 4.02 | — |
Spinal canal | 8.46 ± 3.97 | — | — | — | 8.38 ± 5.79 |
To make a performance assessment of the proposed segmentation method, we compared the average processing time measured from all scans for segmentation of the lungs, trachea/main bronchi, and spinal canal with the methods proposed in other studies [
Average processing time (minutes) of the methods for a 512 × 512 × 100 CT scan.
OARs | Method | ||||
---|---|---|---|---|---|
Our method | de Nunzio et al. [ |
Yim and Hong [ |
Wang et al. [ |
Banik et al. [ | |
Lungs | 3.8 | 5.0 | 6.3 | 8.9 | — |
Trachea/main bronchi | 1.5 | 1.8 | 1.6 | 1.5 | — |
Spinal canal | 4.4 | — | — | — | 9.3 |
In this study, we suggest a fully automated method to segment the lungs, trachea/main bronchi, and spinal canal from CT scans of thorax intended for use in RTP. For this purpose, we implemented software that performs three processes. In the first process, the body region of the patient was segmented by elimination of the background. The lungs and trachea/main bronchi were segmented in the second process and finally, the spinal canal was segmented. Within these processes, a new algorithm for inclusion of excluded pathological areas into the segmented lungs, a modified version of the fuzzy segmentation by morphological reconstruction for spinal canal segmentation, and the well
Comparison of our method with the gold standard using the LIDC data reveals that the proposed method properly reproduces the manual segmentations, similar to other methods (Tables
As shown in Table
Our modified version of the fuzzy segmentation by morphological reconstruction has achieved comparable results to the one presented in the aforementioned study [
The use of our method in RTP may have potential implications. It may improve consistency and concordance in delineation, which is a critical part of RTP. It may also assist in accelerating the clinical workflow through shortening the time of contouring, which is highly desirable in busy clinics.
In conclusion, our proposed method achieved favorable results in patients with lung cancer. This very concise and effective method can avoid heavy computational load and might offer expedited segmentation that can be used in RTP, despite the need for further studies supporting its utilization.
The authors declare that that there is no conflict of interests regarding the publication of this paper.