Volume 70, Issue S2 pp. 37-47

ORIGINAL ARTICLE

Open Access

Clinical validation of a semi-automated segmentation algorithm for target volume definition on planning CT and CBCT in stereotactic body radiotherapy (SBRT) for peripheral lung lesions

Ahmed Allam Mohamed MBBCh, MSc, MD,

Corresponding Author

Ahmed Allam Mohamed MBBCh, MSc, MD

[email protected]

orcid.org/0000-0003-0650-3576

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Correspondence

Ahmed Allam Mohamed, Pauwelsstraße 30, 52072, Aachen, Germany; Tel: +49 241 8089261; Fax: +49 241 8082543; E-mail: [email protected]

Search for more papers by this author

Kathrin Risse MPE, PhD,

Kathrin Risse MPE, PhD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Laura Schmitz MSc,

Laura Schmitz MSc

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Marsha Schlenter MBBCh, MD,

Marsha Schlenter MBBCh, MD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Ahmed Chughtai MBBCh, PhD,

Ahmed Chughtai MBBCh, PhD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Maria Ivanciu MBBCh,

Maria Ivanciu MBBCh

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Michael J. Eble MBBCh, MD,

Michael J. Eble MBBCh, MD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Ahmed Allam Mohamed MBBCh, MSc, MD,

Corresponding Author

Ahmed Allam Mohamed MBBCh, MSc, MD

[email protected]

orcid.org/0000-0003-0650-3576

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Correspondence

Ahmed Allam Mohamed, Pauwelsstraße 30, 52072, Aachen, Germany; Tel: +49 241 8089261; Fax: +49 241 8082543; E-mail: [email protected]

Search for more papers by this author

Kathrin Risse MPE, PhD,

Kathrin Risse MPE, PhD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Laura Schmitz MSc,

Laura Schmitz MSc

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Marsha Schlenter MBBCh, MD,

Marsha Schlenter MBBCh, MD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Ahmed Chughtai MBBCh, PhD,

Ahmed Chughtai MBBCh, PhD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Maria Ivanciu MBBCh,

Maria Ivanciu MBBCh

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

Michael J. Eble MBBCh, MD,

Michael J. Eble MBBCh, MD

Department of Radiation Oncology, RWTH Aachen University, Aachen, Germany

Center for Integrated Oncology Aachen, Bonn, Cologne and Duesseldorf (CIO ABCD), Aachen, Germany

Search for more papers by this author

First published: 24 November 2022

https://doi.org/10.1002/jmrs.637

Share a link

Email
Wechat
Bluesky

Abstract

Introduction

Stereotactic body radiotherapy (SBRT) is an ablative method for lung malignancies. Here, the definition of the gross target volume (GTV) is subject to interobserver variation. In this study, we aimed to evaluate the interobserver variability during SBRT and its dosimetric impact, as well as to introduce a semi-automated delineation tool for both planning computer tomography (P-CT) and cone beam CT (CBCT) to help to standardise GTV delineation and adaptive volume-change registration.

Methods

The interobserver variation of GTV manual contours from five physicians was analysed in 15 patients after lung SBRT on free breathing (FB) P-CT (n = 15) and CBCT (n = 90) before and after each fraction. The dosimetric impact from interobserver variations of GTV based on the original treatment plan was analysed. Next, the accuracy of an in-house easy-to-use semi-automated-segmentation algorithm for pulmonary lesions was compared with gold standard contours in FB P-CT and CBCT, as well as 4D P-CT of additional 10 patients.

Results

The interobserver variability in manual contours resulted in violations of dose coverage of the planning target volume (PTV), which, in turn, resulted in compromised tumour control probability in contours from four physicians. The validation of the semi-automated delineation algorithm using thorax phantom led to a highly reliable accuracy in defining GTVs. Comparing the unsupervised auto-contours with the gold standard delineation revealed high equal high concordance for FB P-CT, 4D P-CT and CBCT, with a DSC of 0.83, 0.76 and 0.8, respectively. The supervised use of the semi-automated delineation tool improved its accuracy, with DSCs of 0.86, 0.86 and 0.8 for FB P-CT, 4D P-CT and CBCT, respectively. The use of the algorithm was associated with a significantly shorter working time. The semi-automated delineation tool can accurately register volume changes in CBCTs.

Conclusion

The segmentation algorithm provides a reliable, standardised and time-saving alternative for manual delineation in lung SBRT in P-CT and CBCT.

Introduction

The role of ablative stereotactic body radiotherapy (SBRT) in early primary lung cancer and pulmonary metastases has evolved rapidly in the last few years. SBRT provides high local control rates comparable to surgical resection while being non-invasive.^1-3 The development of SBRT was based on several advancement milestones in medical imaging and radiation treatment planning. Image-based planning and image-guided treatment delivery to detect and correct tumour motion allowed an improved therapeutic window through the highly conformal spatial dose distribution with an ablative dose to the target and rapid dose fall-off outside the target to minimise the radiation dose to the adjacent tissues. However, it is challenging in SBRT to develop a standardised and objective method for precise target volume definition, which will impede unnecessary volume expansion of the planning target volume (PTV) and a higher risk of geometric miss.⁴ The manual delineation could be time-consuming and highly prone to subjective differences in target delineation.

Multiple solutions for a standardised and more accurate volume delineation have been proposed: adding different imaging modalities, using standard contouring protocols, and the use of an automated contouring process were introduced.⁵ Several automated segmentation methods have been described, such as using voxel changes in identifying the edge or graph cuts, atlas-based automatic segmentation, and most recently, machine learning.⁶

At the same time, the application of cone beam computer tomography (CBCT) in radiation therapy enabled image guidance and, therefore, led to a marked improvement in high-precision radiotherapy. Although CBCT images provide the best real-time body information before irradiation, they are adversely affected by noise and artefacts, which degrade the image quality.⁷ In some circumstances, the image quality is so low that CBCT images cannot be used directly to re-evaluate the delineation of the tumour lesions and replan the dose distribution in adaptive radiotherapy.

This study aimed to validate the use of an easily applicable delineation algorithm for planning CT (P-CT) and CBCT by comparing the interobserver variability from manual and automated contouring of the gross target volume (GTV). Performing a pre-fractional and post-fractional CBCT, the intrafraction and interfraction variation of the GTV volume could be traced. Therefore, we investigated the potential role of automated delineation in standardisation and adaptive target volume delineation.

Material and Methods

Data acquisition

Free breathing (FB) P-CT and CBCTs from 15 patients with a total of 18 peripheral pulmonary lesions (three patients had two lesions) and four-dimensional (4D) P-CT from 10 patients with 10 pulmonary lesions who received SBRT for lung cancer or pulmonary metastases in the department of radiation oncology, RWTH Aachen university medical centre, were included in the current study (Table 1). Patients were selected according to the prescribed dose and the location of the lesion in the lung, being peripheral and not directly adjacent to surrounding structures like ribs or mediastinum. All lesions were analysed independently from each other. All patients were treated with a total dose of 45 Gy in three fractions of 15 Gy prescribed to the 67% isodose on alternative days, using a volumetric modulated arc therapy after signing informed consent for the treatment. Patients were immobilised using evacuated bags and transparent plastic films (BodyFIX 14 BlueBAG; BodyFIX Cover Sheet, ITV, Innsbruck, Austria). The breathing motion was limited passively using an abdominal press (BodyFIX Diaphragm Control ITV, Innsbruck, Austria). The initial free breathing 3-dimensional P-CT in 15 patients was performed on a 16-slice CT scanner (Brilliance CT Big Bore Oncology, Philips Medical Systems Inc., Cleveland, Ohio, USA) using 120 kV, 146 mAs, a pitch of 0.813 and a slice sickness of 3 mm (reconstructed pixel size 0.98 mm × 0.98 mm). During the simulation, the tumour motion was assessed using kV-fluoroscopy to ensure a motion-range ≤5 mm. Before and just after each treatment fraction, 3D CBCT images (XVI, Elekta, Stockholm, Sweden) were acquired using thorax imaging optimised conditions (120 kV and 264 mAs), resulting in a total of 90 CBCTs. The reconstructed pixel size was 1 × 1 mm with a slice thickness of 3 mm. The reconstructed CBCT images were automatically co-registered with the planning CT images using the bone registration algorithm within a clip-box. In a second step, manual registration was employed by aligning the lung lesion visualised in the CBCT centrally to the PTV from reference planning CT.

Table 1. Patients characteristics, FB P-CT (free breathing planning CT), 4D P-CT (four-dimensional planning CT).

Patient	Diagnosis	Histopathology	Staging	Site	No. of lesions	size of lesion-lesions cm³
1 (FB P-CT)	Lung cancer	Sq. cell Ca.	cT1 cN0 cM0	Upper lobe, right	1	1.726
2 (FB P-CT)	Lung cancer	NA (PET avid)	cT3 cN0 cM0	Upper lobe, left	2	18–6.5
3 FB P-CT	Lung cancer	NA (PET avid)		Middle lobe, right	1	6.2
4 FB P-CT	Metastasis	Oropharynx, Sq. cell. Ca.	IV	Upper lobe, left	1	11.42
5 FB P-CT	lung cancer	NA (PET avid)	cT2 cN0 cM0	Lower lobe, right	1	1.38
6 FB-CT	Metastasis	Oropharynx, Sq cell. Ca	IV	Upper lobe, right	2	3.4–1.4
7 (FB P-CT)	lung cancer	Adenocarcinoma	cT1 cN0 cM0	Lower lobe, left- pleural contact	1	3
8 (FB P-CT)	Lung cancer	Adenocarcinoma	cT1 cN0 cM0	Upper lobe, right	1	2.56
9 (FB P-CT)	Metastasis	Melanoma	IV	Lower lobe, left	1	2.09
10 (FB P-CT)	Lung cancer	Adenocarcinoma	cT2 cN0 cM0	Lower lobe, left	1	20.49
11 (FB P-CT)	Lung cancer	NA (PET avid)	cT1 cN0 cM0	Middle lobe, right	1	3.28
12 (FB P-CT)	Lung cancer	Sq. cell Ca.	cT1 cN0 cM0	Upper lobe, left	1	2.471
13 (FB P-CT)	Lung cancer	Adenocarcinoma	cT2 cN0 cM0	Upper lobe, right	2	15–1.97
14 (FB P-CT)	Lung cancer	Adenocarcinoma	cT3 cN0 cM0	Middle lobe, right	1	3.83
15 (FB P-CT)	Metastasis	Melanome	IV	Middle lobe, right	1	3.21
16 (4D P-CT)	Lung cancer	Adenocarcinoma	rcT1 cN0 cM0	Upper lobe, right	1	1.65
17 (4D P-CT)	Lung cancer	Adenocarcinoma	cT2 cN0 cM0	Upper lobe, right	1	13.76
18 (4D P-CT)	Lung cancer	Adenocarcinoma	cT2 cN2 cM0	Middle lobe, right	1	13.1
19 (4D P-CT)	Lung cancer	NA (PET avid)	cT1 cN0 cM0	Upper lobe, left	1	3.65
20 (4D P-CT)	Lung cancer	NA (PET avid)	cT1 cN0 cM0	Upper lobe, left	1	2.6
21 (4D P-CT)	Metastasis	Rectal cancer, Adenocarcinoma	IV	Middle lobe, right	1	3.46
22(4D P-CT))	Lung cancer	Adenocarcinoma	cT1 cN0 cM0	Upper lobe, right	1	13.76
23 (4D P-CT))	Metastasis	Endometrial cancer, Adenocarcinoma	IV	Lower lobe, right	1	2.97
24 (4D P-CT)	Metastasis	Oropharynx, Sq. cell. Ca.	IV	Upper lobe, right	1	1.5
25 (4D P-CT)	Metastasis	Skin, Sq. cell. Ca.	IV	Lower lobe, left	1	2.36

All CBCTs were sent back to the treatment planning system (TPS) Pinnacle³ (V.14.0; Philips Healthcare, Amsterdam, Netherlands), as the CBCTs were reconstructed in the P-CT image coordinates with corrections applied prior to export from the treatment machine, the CBCT images were correctly registered with the P-CT on import to replicate the delivered treatment field position.

For 4D P-CT, the reconstructed CT slices were binned into 10 phases. The Sentinel 4DCT (CRAD, Uppsala, Sweden) was used for motion tracking during the acquisition of the 4D P-CT. The maximum intensity projection CT (MIP) was used to delineate the internal target volume (ITV). FB P-CT, 4D P-CT and CBCT were anonymized for the purpose of the analysis. The current analysis was carried out in compliance with the regulation of the local ethical committee (Faculty of Medicine, RWTH Aachen University), and separate ethical approval was not required for the study.

The semi-automated delineation algorithm

An easy-to-use algorithm was implemented in the treatment planning system Pinnacle³ using the internal automated threshold delineation function and the region growing tool. The function requires the definition of a lower auto-contour threshold value in Hounsfield units (HU) to separate tumour from healthy tissue. The developed algorithm calculates this lower threshold automatically and can be applied to CT images and CBCT images without any change of parameters.

In the first step, the user must define a region of interest (ROI) using the bounding box toolbox of Pinnacle³ (Fig. S1). Within the ROI, the mean HU (meanHUs) and standard deviation (devHUs) are used to calculate the lower threshold based on eq. 1.

lowerthreshold = meanHUs + \frac{devHUs}{2.5}

()

This threshold was then used by the region growing tool, which needs one manual seed point within the tumour lesion. After setting this seed point, the tumour was automatically delineated.

The algorithm was validated by comparing the known volumes from two tumour models in a thorax phantom (Dynamic Thorax Phantom, CIRS, Norfolk, USA) with the volumes defined by the semi-automated delineation algorithm for both imaging modalities; CT and CBCT. The phantom consists of a simplified human thorax geometry constructed of tissue and lung equivalent epoxy materials. The tumour models are water-filled spheres mounted on a rod, both consisting of Plastic Water® DT. The real tumour volumes were 10.12 and 2.90 cm³. The volumes delineated with the algorithm were: 9.94 ± 0.18 cm³ and 2.71 ± 0.07 cm³ for CT images 10.64 ± 0.32 cm³ and 2.97 ± 0.13 cm³ for CBCT images. The variation in volumes corresponded to a radius of <0.4 mm, which is below the pixel size within the data sets. The semi-automatic delineation of both tumour models was repeated 10 times to investigate the influence of manually defining the size of the bounding box. Accordingly, the variation of the delineated volume has a reproducibility of 4%, which corresponds to a variation in the radius of 1.4% (Tables S1 and S2).

Contour generation

Manual delineation of GTVs for all 18 lesions was generated within all data sets from FB P-CTs and CBCTs and ITV of 10 lesions using only the data sets from 4D P-CT (generated using MIP CTs) by an expert with experience of more than 20 years. These delineations were considered ‘gold standard labels’ for further evaluation. Next, five radiation oncologists with an experience range of 3–11 years manually delineated the GTV in each FB P-CT and CBCT image data set by adopting the delineation protocol from RTOG 0813⁸ to reduce the interobserver variability and being blinded from the other contours. The delineation was performed using a standardised lung window level (window width 1601 HU, window level −300 HU). A mediastinal window (window width 401 HU, window level 800 HU) was also able to be used to ensure the exclusion of adjacent benign structures (vessel or chest wall) or atelectasis, these contours were named ‘manual contours’.

The unsupervised contours using the semi-automated delineation tool were named ‘unsupervised auto-contours’ and were generated by K.R and L.S.

Three months later, two radiation oncologists then evaluated the exact copy of the unsupervised semi-auto-contours. The radiation oncologists were allowed during the evaluation process to edit these contours to exclude any nearby vessels or bronchi. These contours were named ‘supervised auto-contours’.

Quantitative evaluation for manual contours and auto-delineation

The manual contours (FB P-CTs, CBCTs and 4D P-CT) from five physicians as well as the unsupervised and supervised auto-contours were evaluated for interobserver variability by assessing the spatial overlap to the gold standard dataset using the following indices:

Coefficient of variation (CoV)⁹: CoV is a descriptive metric for the range of the volume variation between different contours; ideally, it should be low as possible.

CoV = \frac{σ}{μ}

()

Where

σ

is the Standard Deviation and

μ

is the mean volume of the contours from all observers for each FB P-CT, 4D P-CT and CBCT.

Dice similarity coefficient (DSC)^{9, 10}: DSC is a statistical metric used to evaluate the accuracy of spatial overlap between two volumes (X and Y as per formula 3), where ‘0’ means no overlap between the two GTVs and ‘1’ means that both GTVs are perfectly overlapping on each other. An spatial agreement between both GTVs > 0.7 is considered acceptable.¹¹

DSC = \frac{2 (X \cap Y)}{(X) + (Y)}

()

The time needed for the delineation: The time necessary for the delineation of GTVs was recorded for all the physicians and the operator to run the auto-delineation algorithm for further evaluation.

Dosimetric and radiobiological treatment plan evaluation

Planning target volume was generated for all 18 lesions in FB P-CTs after manual GTVs delineation, expanding the GTV 1 cm craniocaudal and 0.5 cm axial. The dose that covers 95% of each PTV (D95-PTV) was evaluated using the dose-volume histogram (DVH) from the original plan for each patient. Furthermore, the tumour control probability (TCP) in each plan for the PTV from each physician was evaluated; the TCP modelling using a linear-quadratic model based on the following formula¹²:

TCP = \frac{1}{1 + {(\frac{D 50}{D})}^{4 γ}}

()

Where D50 is the dose needed to have 50% tumour control probability and γ is the slope of the sigmoid-shaped dose–response curve at D50.

D 50 : 49.2 Gy & α / β : 10 & γ : 1 .

Finally, the auto-contours volume and the dose covering 100% of GTV (D100-GTV) were calculated for all CBCT contours to evaluate the interfraction and intrafraction tumour volume and dose variation.

Statistical analysis

Correlation analysis with Spearman Rank Correlation was calculated using Free Statistics Software (v1.2.1).¹³ A one-sided paired T-test was used to evaluate the variation between D95-PTV, TCP-PTV, median CoV among the three fractions in CBCTs, and the time needed for contouring. Graphs were generated using Excel (Microsoft, CA, USA) and SPSS version 25 (IBM, New York, NY, USA).

Results

Interobserver variability in GTV delineation from P-CT and CBCTs

After the manual delineation of GTVs in FB P-CT, the mean and standard deviation of the CoV from the five physicians among the eighteen lesions was 0.17 ± 0.06 (Fig. S2a). Four patients (number 5, 7, 9 and 12) showed a higher CoV (>0.2), indicating a higher interobserver variation. Despite frequent higher CoV in smaller lesions, the statistical analysis revealed no correlation between the size of the lesions and the CoV (rho: −0.261 and 2-sided P-value: 0.347) (Fig. S2b). The mean CoV and standard deviation for manual delineations of lesions in CBCTs was 0.21 ± 0.1 (Fraction ‘Fr’ 1: 0.17 ± 0.08, Fr2: 0.21 ± 0.1, Fr3: 0.24 ± 0.11). The CoV for manual delineation of the GTVs in CBCTs was significantly increased in the second and third radiation fraction, compared to the first fraction, with p-value 0.005 and 0.002, respectively (Fig. S2c,d), indicating a poorer interobserver agreement during the course of treatment.

The mean DSC, which demonstrates the variation in the spatial overlap between contours from all five physicians and the ground truth contours in FB P-CTs for the 18 lesions, was 0.83 (Fig. S3a). The analogous mean DSC after manual contouring in CBCTs was 0.80 (Fig. S3b). The mean DSC for P-CT and CBCT will be used later to validate the auto-contours (unsupervised and supervised) and will be named ‘reference manual’.

Impact of interobserver variability on PTV size, D95-PTV and TCP-PTV

The resulting PTVs from manual GTV re-delineation in FB P-CTs were used to evaluate the corresponding dose-volume load based on the implementation of the new target volume delineation upon the original dose distribution of the initial radiation treatment plan for each patient (PTV-irradiated). Overall, the variation in manual delineations of the GTV among the 5 physicians led to variations in PTV size with an increase in mean PTV size from manual delineations in 12 patients, a decrease in two patients and a similar in one patient compared to PTV-irradiated (Table 2, Fig. 1c). These variations were translated into significantly lower D95-PTV than the original radiation plan for each patient for all manual contours (Fig. 1a). Consequently, these lower D95-PTV resulted in substantially lower TCP values for the re-planned dose application (Fig. 1b).

Table 2. The size of the PTVs from 15 FB P-CTs (free breathing planning CTs) for all five observers in cm³.

	Pt1	Pt2	Pt3	Pt4	Pt5	Pt6	Pt7	Pt8	Pt9	Pt10	Pt11	Pt12	Pt13	Pt14	Pt15
PTV Physician 1	13.61	105.2	37.26	31.44	13.34	38.9	24.38	22.7	15.26	91.3	20.87	18.9	88.9	25	21
PTV Physician 2	12.9	118	31.52	30.35	15.36	37	20.75	24.18	18.9	97.6	20.6	16.64	66.57	23.81	21.3
PTV Physician 3	12.75	130.4	34.83	29.55	16.37	42.7	31.62	26.84	16.41	79.42	21.32	17.56	97.08	26.6	24.2
PTV Physician 4	13.24	108.31	30.11	27.15	14.45	43.5	22.9	24.01	22.94	79.94	24.36	24.12	91.3	28.89	22.5
PTV Physician 5	13.35	94.83	31.33	22.06	10.96	32.63	19.8	19.9	13.9	65.35	20.86	16.92	71	23.23	20.3
Mean (SD)	13.17 (±0.35)	111.35 (±13.48)	33.01 (±2.95)	28.11 (±2.08)	14.1 (±2.08)	38.95 (±4.43)	23.89 (±4.68)	23.57 (±2.54)	17.48 (±3.56)	82.77 (±12.40)	21.6 (±1.56)	18.83 (±3.08)	82.97 (±13.38)	25.51 (±2.29)	21.86 (±1.53)
PTV-irradiated	10.11	80.57	39.47	20.79	12.45	27.46	29.95	23.33	12.477	70.6	18.1	17.45	50.64	18.48	20.3

Details are in the caption following the image — **Figure 1**
Open in figure viewer PowerPoint

(a) Boxplot generated from the values of the dose that covers 95% of each PTV (D95-PTV) from the original plan of the 15 free breathing (FB) P-CTs and contours from the five physicians (manual contours) based on the original radiation plan for each patient (**P-value: <0.005, ***P-value <0.0005, t-test). (b) Boxplot generated from the tumour control probability (TCP) based on PTV coverage for each physician based on the original radiation and compared with the TCP from the original contours (P-value: ** <0.005 & *** <0.0005, t-test). (c) Colourwash dose distribution of the original radiotherapy plan in patient number 10 in axial and coronal view to explain the underdosing of the PTVs from different observers. The original GTV in blue and the original PTV in purple, GTVs and PTVs of the five physicians are red and sky blue.

The accuracy of the semi-automated delineation tool

The median DSC for the unsupervised auto-contours in FB P-CT was 0.83, and the median of DSC for the supervised auto-contours was 0.86 (Fig. 2a). Further, the median of DSC for unsupervised auto-contours and supervised auto-contours in 4D P-CTs were 0.76 and 0.86, respectively (Fig. 2a). Also, the median of DSC for the unsupervised auto-contours and supervised auto-contours in CBCTs were 0.80 (Fig. 2b), indicating an acceptable accuracy of the delineation algorithm in both P-CT and CBCT.

Furthermore, the time consumed for the algorithm's delineation was significantly and stably lower (0.81 ± 0.2 min) than all five physicians' time (1.3 ± 1, 1.1 ± 0.52, 2.66 ± 1.89, 1.46 ± 0.5 and 1.27 ± 1.12 min) (Fig. 2c).

The tracking of the interfraction and intrafraction GTV volume changes

Based on the pre-radiation (pre-RT) images for each fraction, the evaluation of the interfraction GTV volume changes during radiotherapy, between fractions 1&2 and fractions 2&3, showed an increase in GTV volume in most patients. While a mean volume increase of 33.4% (range 6.3–72.1%) was observed in 11 patients, only four patients revealed a lesser pronounced decrease in tumour volume (4.6%, range 2.7–6.7%). The ratio of patients with volume increase and decrease was equal between unsupervised auto-and gold standard contouring; however, the mean value in volume changes was higher after semi-automatic segmentation (fraction 1–2 11.6% vs. 18.4%; fraction 2–3 18.4% vs. 23.25%) (Fig. 3a). The difference in GTV volume data between semi-automated segmentation and gold standard delineation was non-significantly different (the mean: 14.6% manual contouring, 20.5% automatic segmentation). The standard deviation was equal for both delineation methods (17.5% and 17.8%) with a broad range of percent volume changes.

The intrafraction GTV volume changes were evaluated using the pre-RT and immediate post-RT CBCTs for the corresponding fraction during radiotherapy. Overall there was an intrafractional volume decrease both in the gold standard and unsupervised automatic delineation (Fig. 3b). semi-automated delineation revealed a decreased GTV volume immediately after SBRT in 4 patients for fraction 1, 5 patients for fraction 2 and 10 patients for fraction 3, and gold standard delineations in 4, 9 and 9 patients, respectively. The standard deviation demonstrated a broader range of values for both contours (7.85 and 14.14%).

Evaluating the impact of the intrafraction tumour volume changes (using the unsupervised auto-contours) on the initial treatment plan resulted in an underdosing of mean D100-GTV of 9.45% ±5.2% in 11 patients within the 1st fraction, 7.95% ±5.94% in 12 patients in the 2nd fraction and 9.31% ±5.63% in 7 patients in the 3rd fraction (Fig. 3c).

Discussion

In the current study, we emphasised the impact of the interobserver variation of GTV delineation in SBRT of pulmonary lesions through the evaluation of its effect on the dose coverage of the target volumes based on the original radiation plan for each patient. Unsurprisingly, we found that the interobserver variation in the GTV delineation (manual contours) in a cohort of 15 patients resulted in a systematic and significant underdosing for the PTVs from all the participating physicians, which significantly lower TCP at least in the contours from four physicians.

The mean CoV among all five observers in FB P-CTs was 0.17 ± 0.06. The interobserver variation was high in contours of patient number 5,7,9 and 12 in P-CTs with CoV > 0.2. While all those four lesions were relatively small (size ≤ 3cm³), the correlation between CoV and the size of the lesions was not statistically significant in contrast to what was previously reported by Altoraji et al.²⁰ The target volume delineation in CBCTs was more challenging, with more frequent high CoV. In particular, we observed a statistically significant higher interobserver variation in the later CBCTs and fractions. We do not have an explanation for such an observation. Still, a possible suggestion may be the increased radiation-induced changes in the lung parenchyma or tumour deformation through the course of the irradiation, which reduces the sharpness of tumour boundaries.

With improvements in the conformal dose application, the impact of the interobserver variability in target volume delineation evolved into a significant issue in radiation oncology.^14-16 The main concerns are missing part of the target volume or involving normal tissue in target volumes. Subsequently, this could result in undermining tumour control, as we emphasised in the current study or increasing toxicities.^{17, 18} Also, interobserver variability could impact the robustness of the radiomics features negatively, and the standardisation of target volume delineation could improve the stability of the prognostic value of the radiomics features.¹⁹

In the current study, we introduced a semi-automated delineation tool for target volume delineation based on the change of HU units to define the tumour boundaries using a function available in the TPS Pinnacle,³ which is easily applicable in the clinical environment within the TPS. This semi-automated delineation tool is a simple approach that uses the HUs in the lesion and the surrounding tissue to define the lower threshold, which makes it more patient-specific than other approaches using a fixed HU Threshold,²¹ therefore, can be used for P-CT datasets and CBCT datasets. The algorithm's accuracy was validated using a lung phantom with known lesion size. Using the algorithm, the variation in the radius of the delineated lesion size on imaging is <0.4 mm, which is smaller than the pixel size of the image datasets. The accuracy and repeatability are in the same order as other automated delineation algorithms.^{21, 22}

For further validation of the semi-automated delineation algorithm, we examined the DSC among the five physicians as human operators in reference to ground truth labels in 15 FB P-CTs and 19 CBCTs. Further, we calculated the mean of all the medians DSCs from all five operators in both FB P-CT and CBCTs to be used as a ‘reference manual’ to validate the contours from the algorithm, which were 0.83 and 0.80, respectively. The median DSC for the auto-contours from the algorithm in FB P-CTs and CBCTs was 0.83 and 0.80, respectively, indicating that auto-contours are clinically reliable even in limited quality images as CBCTs.

We experienced some shortcomings in the delineation algorithm, for example the inclusion of an adjacent small vessel within the GTV (Fig. S4). To overcome such a problem, we allowed a modification of the auto-contours by the agreement of two observers 3 months after the manual delineation to ensure less influence of the personal judgement based on the previous contouring process. This improved the median DSCs in P-CT, enhanced the lower values and reduced the range in DSCs in CBCTs.

Further, we examined the semi-automated delineation tool using 4D P-CT. For this purpose, we included a second cohort of 10 4D P-CTs, and the delineation was carried out in MIP CTs. Here, the DSCs for the auto-contours and modified auto-contours were 0.76 and 0.86, respectively.

Nevertheless, we believe that applying an automated delineation method should be simple and does not consume time besides the accuracy to improve its feasibility in routine application. Indeed, we showed that the application of the current semi-automated delineation tool significantly reduced the time needed for the delineation process in a stable manner compared to the manual delineation adding to its value.

Another possible application of the semi-automated delineation tool is tracking the interfraction and intrafraction changes and deformations. The tumour interfraction, intrafraction deformation and size variation have been widely discussed in lung SBRT.^{21, 23} Theoretically, a reduction in the size would not negatively impact the dosimetric outcome. In contrast, the increase in the size of the irradiated lesion compared to the initial P-CT may adversely affect the dose distribution in PTV.^{24, 25} Another study retrospectively observed the interfraction tumour volume change in twenty-two patients. It concluded a correlation between a marked reduction in the size during the SBRT in early lung cancer and worse clinical outcomes, progression-free survival and overall survival.²⁶ Repeated CBCT before and after each fraction has been proposed to address such a problem as a tool for the target volume monitoring in lung SBRT.²⁷ The addition of an automated method to delineate the target volume would help to figure out the magnitude of intrafraction and interfraction tumour deformations and the necessity to repeat the planning in case of undesired dosimetric violations, known as adaptive radiation therapy (ART).²⁸ In the current study, we tested the semi-automated delineation tool in detecting such variation over repeated CBCTs before and after each fraction and the dosimetric difference that resulted from such variation on D100-GTV.

Indeed, we observed a deformation over the multiple CBCTs with an increase in GTV volume in 60% of the patients, a reduction in the volume in the subsequent CBCTs in 13.4% of the patients and mixed variation in the size (increase and decrease) in 26.66% of the patients. We cannot explain with certainty these changes in the volume of GTVs in CBCTs, but a possible contribution to such changes could be the irregularity of breathing patterns during the radiation course.²⁹ Still, the automated contouring for online and speed delineation of the GTV in CBCT would provide a tool for detecting intrafraction and interfraction volume changes that could be better compensated in the calculation process of PTV margins.

Limitations of the Study

There are some limitations of the current study that should be openly addressed. the semi-automated delineation tool can be applied mainly for the peripheral lesions without large contact with the chest wall or mediastinum, as it cannot accurately differentiate between the tumour boundaries and thorax wall or mediastinum. This would reduce the applicability of the tool. The necessity to revise the delineation due to the possible inclusion of nearby vessel is another limitation of the tool. Finally, the need to define ROI to apply the tool would make the tool semi-automated rather than a fully automated tool.

Conclusion

Taken together, the supervised use of the aforementioned semi-automated delineation tool provides a clinically reliable and standardised target volume delineation tool in both P-CT and CBCT in peripheral lung SBRT. It is user-friendly, saves time and can detect any deformation or volume changes of the targets over the radiation course, allowing the chance for adaptive consideration of the target volume.

Funding information

Open access funding enabled and organised by Projekt DEAL.

Conflict of interest

All authors declare that they have no competing interests.

Acknowledgement

Open Access funding enabled and organized by Projekt DEAL.

Open Research

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Supporting Information

References

1Chang JY, Senan S, Paul MA, et al. Stereotactic ablative radiotherapy versus lobectomy for operable stage I non-small-cell lung cancer: a pooled analysis of two randomised trials. Lancet Oncol 2015; 16: 630–7.
10.1016/S1470-2045(15)70168-3
PubMed Web of Science® Google Scholar
2Guckenberger M, Andratschke N, Alheit H, et al. Definition of stereotactic body radiotherapy: Principles and practice for the treatment of stage I non-small cell lung cancer. Strahlentherapie und Onkol 2014; 190: 26–33.
10.1007/s00066-013-0450-y
CAS PubMed Web of Science® Google Scholar
3Klement RJ, Hoerner-Rieber J, Adebahr S, et al. Stereotactic body radiotherapy (SBRT) for multiple pulmonary oligometastases: analysis of number and timing of repeat SBRT as impact factors on treatment safety and efficacy. Radiother Oncol 2018; 127: 246–52.
10.1016/j.radonc.2018.02.016
CAS PubMed Web of Science® Google Scholar
4Njeh CF. Tumor delineation: The weakest link in the search for accuracy in radiotherapy. J Med Phys 2008; 33: 136–40.
10.4103/0971-6203.44472
CAS PubMed Google Scholar
5Vinod SK, Min M, Jameson MG, Holloway LC. A review of interventions to reduce inter-observer variability in volume delineation in radiation oncology. J Med Imaging Radiat Oncol 2016; 60: 393–406.
10.1111/1754-9485.12462
PubMed Web of Science® Google Scholar
6Sharp G, Fritscher KD, Pekar V, et al. Vision 20/20: Perspectives on automated image segmentation for radiotherapy. Med Phys 2014; 41: 050902.
10.1118/1.4871620
PubMed Web of Science® Google Scholar
7Schulze R, Heil U, Groß D, et al. Artefacts in CBCT: a review. Dentomaxillofacial Radiol 2011; 40: 265–73.
10.1259/dmfr/30642039
CAS PubMed Web of Science® Google Scholar
8Bezjak A, Paulus R, Gaspar LE, et al. Safety and efficacy of a five-fraction stereotactic body radiotherapy schedule for centrally located non–small-cell lung cancer: NRG Oncology/RTOG 0813 trial. J Clin Oncol 2019; 37: 1316–25.
10.1200/JCO.18.00622
CAS PubMed Web of Science® Google Scholar
9Fotina I, Lütgendorf-Caucig C, Stock M, Pötter R, Georg D. Critical discussion of evaluation parameters for inter-observer variability in target definition for radiation therapy. Strahlenther Onkol 2012; 188: 160–7.
10.1007/s00066-011-0027-6
CAS PubMed Web of Science® Google Scholar
10Dice LR. Measures of the amount of ecologic association between species. Ecology 1945; 26: 297–302.
10.2307/1932409
Web of Science® Google Scholar
11Thada V, Jaglan V. Comparison of Jaccard, dice, cosine similarity coefficient to find best fitness value for web retrieved documents using genetic algorithm. Int J Innov Eng Technol 2013; 2: 202–5.
Google Scholar
12Martel MK, Ten Haken RK, Hazuka MB, et al. Estimation of tumor control probability model parameters from 3-D dose distributions of non-small cell lung cancer patients. Lung Cancer 1999; 24: 31–7.
10.1016/S0169-5002(99)00019-7
CAS PubMed Web of Science® Google Scholar
13 Spearman Rank Correlation - Free Statistics and Forecasting Software (Calculators) v.1.2.1, [Accessed 2020 August 30] Available from: https://www.wessa.net/rwasp_spearman.wasp#output
Google Scholar
14Gkika E, Tanadini-Lang S, Kirste S, et al. Interobserver variability in target volume delineation of hepatocellular carcinoma: an analysis of the working group “Stereotactic Radiotherapy” of the German Society for Radiation Oncology (DEGRO). Strahlentherapie und Onkol 2017; 193: 823–30.
10.1007/s00066-017-1177-y
CAS PubMed Web of Science® Google Scholar
15Tsang Y, Hoskin P, Spezi E, et al. Assessment of contour variability in target volumes and organs at risk in lung cancer radiotherapy. Tech Innov Patient Support Radiat Oncol 2019; 10: 8–12.
10.1016/j.tipsro.2019.05.001
PubMed Google Scholar
16Persson GF, Nygaard DE, Hollensen C, et al. Interobserver delineation variation in lung tumour stereotactic body radiotherapy. Br J Radiol 2012; 85: e654–60.
10.1259/bjr/76424694
CAS PubMed Web of Science® Google Scholar
17Eminowicz G, Rompokos V, Stacey C, McCormack M. The dosimetric impact of target volume delineation variation for cervical cancer radiotherapy. Radiother Oncol 2016; 120: 493–9.
10.1016/j.radonc.2016.04.028
PubMed Web of Science® Google Scholar
18Lo AC, Liu M, Chan E, et al. The impact of peer review of volume delineation in stereotactic body radiation therapy planning for primary lung cancer: a Multicenter Quality Assurance Study. J Thorac Oncol 2014; 9: 527–33.
10.1097/JTO.0000000000000119
PubMed Web of Science® Google Scholar
19Kothari G, Woon B, Patrick CJ, et al. The impact of inter-observer variation in delineation on robustness of radiomics features in non-small cell lung cancer. Sci Rep 2022; 12: 1–11.
10.1038/s41598-022-16520-9
PubMed Web of Science® Google Scholar
20Altorjai G, Fotina I, Lütgendorf-Caucig C, et al. Cone-beam CT-based delineation of stereotactic lung targets: the influence of image modality and target size on interobserver variability. Int J Radiat Oncol Biol Phys 2012; 82: e265–72.
10.1016/j.ijrobp.2011.03.042
PubMed Google Scholar
21Tatekawa K, Iwata H, Kawaguchi T, et al. Changes in volume of stage I non-small-cell lung cancer during stereotactic body radiotherapy. Radiat Oncol 2014; 9: 8.
10.1186/1748-717X-9-8
PubMed Web of Science® Google Scholar
22Schaefer A, Kremp S, Hellwig D, Rübe C, Kirsch CM, Nestle U. A contrast-oriented algorithm for FDG-PET-based delineation of tumour volumes for the radiotherapy of lung cancer: derivation from phantom measurements and validation in patient data. Eur J Nucl Med Mol Imaging 2008; 35: 1989–99.
10.1007/s00259-008-0875-1
PubMed Web of Science® Google Scholar
23Salamekh S, Rong Y, Ayan AS, et al. Inter-fraction tumor volume response during lung stereotactic body radiation therapy correlated to patient variables. PLoS One 2016; 11: e0153245.
10.1371/journal.pone.0153245
PubMed Web of Science® Google Scholar
24Wu Q, Liu C, Devpura S, Ajlouni M, Movsas B, Chetty I. Accumulated dose estimation in lung stereotactic body radiation therapy using deformable image registration algorithm. Med Phys 2017; 44: 3222–3.
Web of Science® Google Scholar
25Jia J, Tian Z, Gu X, Yan H, Jia X, Jiang S. Investigating the dosimetric effect of inter-fraction deformation in lung cancer stereotactic body radiotherapy (SBRT). Int J Cancer Ther Oncol 2014; 2: 020225.
10.14319/ijcto.0202.25
Google Scholar
26Pathak RS, Tibdewal AR, Laskar SG, Chaudhari S, Agarwal JP. Volume changes with stereotactic body radiation therapy in early lung cancer: time trends and outcomes. J Med Imaging Radiat Oncol 2016; 60: 661–7.
10.1111/1754-9485.12476
PubMed Web of Science® Google Scholar
27Hou Y, Lee S, Agrawal V, et al. Inter-scan and inter-observer tumour volume delineation variability on cone beam computed tomography in patients treated with stereotactic body radiation therapy for early-stage non-small cell lung cancer. J Med Imaging Radiat Oncol 2017; 61: 93–8.
10.1111/1754-9485.12537
PubMed Web of Science® Google Scholar
28Mao W, Riess J, Kim J, et al. Evaluation of auto-contouring and dose distributions for online adaptive radiation therapy of patients with locally advanced lung cancers. Pract Radiat Oncol 2022; 12: e329–38.
10.1016/j.prro.2021.12.017
PubMed Web of Science® Google Scholar
29Clements N, Kron T, Franich R, et al. The effect of irregular breathing patterns on internal target volumes in four-dimensional CT and cone-beam CT images in the context of stereotactic lung radiotherapy. Med Phys 2013; 40: 021904.
10.1118/1.4773310
CAS PubMed Web of Science® Google Scholar

Volume70, IssueS2

Special Issue:Horizon scanning: The impact of technology and integration on medical radiation sciences profession

April 2023

Pages 37-47

This article also appears in:

Clinical validation of a semi-automated segmentation algorithm for target volume definition on planning CT and CBCT in stereotactic body radiotherapy (SBRT) for peripheral lung lesions