Enhancing the Contouring Efficiency for Head and Neck Cancer Radiotherapy Using Atlas-based Auto-segmentation and Scripting

Abstract

Background/Aim

Intensity-modulated radiation therapy can deliver a highly conformal dose to a target while minimizing the dose to the organs at risk (OARs). Delineating the contours of OARs is time-consuming, and various automatic contouring software programs have been employed to reduce the delineation time. However, some software operations are manual, and further reduction in time is possible. This study aimed to automate running atlas-based auto-segmentation (ABAS) and software operations using a scripting function, thereby reducing work time.

Materials and Methods

Dice coefficient and Hausdorff distance were used to determine geometric accuracy. The manual delineation, automatic delineation, and modification times were measured. While modifying the contours, the degree of subjective correction was rated on a four-point scale.

Results

The model exhibited generally good geometric accuracy. However, some OARs, such as the chiasm, optic nerve, retina, lens, and brain require improvement. The average contour delineation time was reduced from 57 to 29 min (p<0.05). The subjective revision degree results indicated that all OARs required minor modifications; only the submandibular gland, thyroid, and esophagus were rated as modified from scratch.

Conclusion

The ABAS model and scripted automation in head and neck cancer reduced the work time and software operations. The time can be further reduced by improving contour accuracy.

Intensity-modulated radiation therapy (IMRT) employed in head and neck cancer (HNC) enables the delivery of a highly conformal dose to a target and minimizes the dose to organs at risk (OARs) by delivering a complex, steep dose distribution (1). Compared with three-dimensional conformal radiation therapy, IMRT significantly reduces the dose to OARs and radiation-related toxicity symptoms, such as difficulty in opening the mouth, impaired chewing, speech, swallowing, and development of dental caries (2-4). Therefore, for high-precision radiotherapy planning, accurate segmentation of the location and size of the OARs is essential. However, time and effort are required to delineate contours because numerous OARs with complex shapes exist in the head and neck region. Previous studies have reported that an average of 2.7-3 hours is required to manually delineate contours (5,6). To reduce this time-consuming procedure, atlas-based auto-segmentation (ABAS) has been increasingly used in clinical practice. ABAS is an algorithm that deforms computed tomography (CT) images and delineates contours using an atlas template that registers the CT images and contours of another patient (7). Qazi et al. reported that the mean Dice similarity coefficient (DSC) between manual contours and ABAS exceeds 0.8 (8). Several studies have reported that ABAS significantly reduces delineation time and improves planning efficiency (5,9,10). While various treatment planning systems (TPS) are available, some do not allow the use of ABAS. The workflow for using ABAS is as follows: send CT images to a third vendor’s auto-contouring software (ACS), perform ABAS in the vendor’s software, transfer the completed structure set to the TPS, and import it into the TPS. Waiting times, particularly in the head and neck areas, are due to the time required for ABAS, which involves many OARs. TPSs and ACSs possess scripting capabilities that can improve this process. This study aimed to evaluate the accuracy of ABAS, construct a script, and compare the time required to modify the contours delineated by scripted ABAS versus manual contouring.

Materials and Methods

CT simulation. CT images of 69 patients with HNC treated with chemoradiotherapy were obtained. Table I shows the patient characteristics. All patients received volumetric modulated arc therapy for HNC between 2021 and 2023 at the Osaka International Cancer Institute while immobilized with an individual nine-pin thermoplastic mask (Civco, Head and shell Type S, Zwanenburg, the Netherlands). CT images were obtained using a GE Dual Energy instrument (Revolution HD; GE Medical Systems, Milwaukee, WI, USA). The parameters for image acquisition were as follows: tube voltages of 80 and 140 kVp and a tube current of 550 mA. The reconstruction parameters were as follows: slice thickness, 2 mm; pixel matrix, 512×512 pixels; field of view, 35 cm. The CT images were transferred to the Eclipse planning system (version 15.6.05; Varian Medical Systems, Palo Alto, CA, USA).

Table I. Patient characteristics.

User-defined contours. The CT images were transferred to the TPS, and radiation oncologists delineated the OARs according to the Radiation Therapy Oncology Group guidelines (11). The following OARs were delineated: the brain, brainstem, chiasm, optic nerve, retina, cochlea, oral cavity, mandible, pharyngeal constrictor muscle (PCM), larynx, parotid gland, submandibular gland, spinal cord, esophagus, and thyroid. The average values were calculated for the left and right sides.

ABAS software. Multi-atlas auto-segmentation algorithms were used for ABAS in the RayStation Research Ver. 10.0.1.52 (Ray Search Laboratories, Stockholm, Sweden). Thirty patients were registered in the atlas model dataset for modeling. ABAS uses the ANAtomically CONstrained Deformation Algorithm (ANACONDA) for image deformation (12). ANACONDA uses intensity-based and anatomical information-based approaches to calculate deformation vectors to achieve the best match between images using an atlas model dataset.

Evaluation of geometric segmentation accuracy. Fifteen patients who were not included in the ABAS model were used to evaluate the geometric accuracy of the manually and automatically delineated contours. Both contours were transferred to MIM Maestro ver.7.3.4 (MIM, MIM Software Inc., Cleveland, OH, USA) for subsequent analyses. The Dice similarity coefficient (DSC) and the Hausdorff distance (HD) were used. The DSC is an overlapping area that measures the degree of overlap between two different volumes and is defined by an equation (13). The HD is the maximum absolute distance between the manually and automatically delineated contours (14).

where A denotes the reference contour; B denotes the contour delineated by the software; a and b denote the points on A and B, respectively; and d (a, b) denotes the Euclidean distance between a and b.

Scripting. A schema of the contour delineation workflow developed in this study is shown in Figure 1. The left side shows the conventional method, which manually delineates the contours, and the right side shows the method that incorporates ACS. Conventionally, CT radiographers transfer the CT images obtained using TPS. A treatment planner manually imported the images and delineated contours. In the newly developed method shown on the right, the CT radiographer transfers the CT images to the ACS. The treatment planner then imports data to the ACS. Thereafter, RayStation’s scripting function in IronPython is used to perform the following operations: choose an atlas, create a structure set of OARs, perform ABAS, and save and transfer contours to the TPS. After the contours are imported to TPS, the treatment planner manually checks and modifies them. The IronPython language is a platform that involves the implementation of Python Language (15) combined with a Microsoft Network framework (16).

Figure 1. A schema of the workflow for delineating contours.

Time evaluation and subjective delineation accuracy. Four radiation oncologists and three radiotherapists manually delineated the contours of twenty-five patients with HNC. Concurrently, a script was used to automatically delineate the contours. Thereafter, the automatically delineated contours were manually modified. The time required to complete each of the aforementioned procedures was recorded. For the time comparison, the Wilcoxon signed-rank test was performed to determine statistical significance using SPSS Statistics (IBM, ver.24, Armonk, NY, USA). When modifying the contours, the respondents subjectively rated each OAR as follows: acceptable (no modification is required), minor change is required (50% or more time savings can be expected), major change is required (0-50% time savings can be expected), and not acceptable (it is faster to delineate from scratch).

Results

The ABAS atlas model was constructed using 30 patients with HNC who were randomly selected from a clinical database. Using this atlas model, the OAR contours were delineated automatically in 15 patients with HNC who were not part of the atlas model. Figure 2 shows examples of the automatically delineated contours. Most OARs were well-delineated, but a slight shift was observed in the contours. For example, in the retinas and optic nerves shown in Figure 2B, the shapes of the delineated contours were satisfactory, but the positions were not well-delineated. In the parotid glands shown in Figure 2C, deviations were observed at the boundaries of the bones, blood vessels, and muscles. In the submandibular glands shown in Figure 2D, the contours were delineated and differed significantly in shape.

Figure 2. The automatically delineated contours and computed tomography scans: (A) sagittal slice, (B) upper axial slice, (C) middle axial slice, (D) lower axial slice. The manual contours are shown in magenta (brain), blue (brainstem), yellow (spinal cord), tan (chiasm), brown (oral cavity), fluorescent green (mandible), pink (larynx), ivory (thyroid), pale orange (PCM), red (esophagus), purple (right optic nerve), sky blue (left optic nerve), gold (right retina), silver (left retina), dark red (right lens), green (right parotid gland), yellow green (left parotid gland), navy blue (left submandibular gland), and cyan (right submandibular gland).

The DSC and HD values were calculated by comparing the contours delineated manually and automatically, as shown in Figure 3. Small and low OARs, such as the chiasm, lens, optic nerve, and retina, tended to have lower DSCs than those of other OARs. The mean DSC of the parotid gland, brainstem, spinal cord, oral cavity, and brain exceeded 0.8, indicating a good contour. The HD values for brains with DSCs greater than 0.9 were longer than those of other OARs. Figure 4 shows an example of the brain contours in slices with high HD values. As shown in Figure 4A and C, the contours between the brain and skull were displaced. Moreover, the contours around the ethmoid sinus were displaced, as shown in Figure 4B. Other OARs were delineated with a mean HD value within 10 mm and considered closely delineated.

Figure 3. Box-and-whisker plots of (A) Dice similarity coefficient (DSC) (upper) and (B) Hausdorff distance (HD) value (bottom) were created for 16 organs at risk comparing automatically and manually delineated contours. For structures with left and right sides, the average value was used. In each box, the central mark represents the median, whereas the ends correspond to the 25th and 75th percentiles. The upper and lower whiskers indicate the maximum and minimum values, respectively.

Figure 4. Comparison of contours between automatic and manual brain delineations with long Hausdorff distance (HD) values. Two patients are included, with upper (A and B) and lower (C and D) sections. (A) Patient A, sagittal slice, (B) patient A axial slice, (B) patient B sagittal slice, (D) patient B, axial slice. The HD values were 39 and 38 mm for patients A and B, respectively. Automatic and manual contours are shown in magenta and yellow, respectively.

Figure 5 compares the manual delineation time, running scripting, and manual modification of the contours. The mean manual delineation time was 57 min and the mean running script and manual modification times were 29 min. Time reduction was significant (p<0.001). Individual differences in delineation times were also reduced.

Figure 5. Box-and-whisker plot comparing the time required for manual contour delineation and manual modification after running ABAS. In each box, the central mark represents the median, the cross mark indicates the average, and the edges correspond to the 25^th and 75^th percentiles. The upper and lower whiskers represent the maximum and minimum values, respectively. The Wilcoxon signed-rank sum test was used for the analyses. p<0.001.

Figure 6 shows the subjective four-grade scale (acceptable, minor change required, major change required, and unacceptable) when modifying the contours that were delineated automatically. All OARs—except the thyroid, submandibular gland, and esophagus—showed good results with a few modifications. The low DSC of the chiasm, lens, optic nerve, and retina did not require many modifications. Brains with high HD values required minimal modifications. Figure 7 shows the contours of the submandibular gland, esophagus, and thyroid, which must be delineated from scratch with a subjective degree of revision. As shown in Figure 7A, the left submandibular gland was broadly delineated. In Figure 7B, the shape of the thyroid did not fit, and the surrounding blood vessels were displaced. The esophagus was delineated differently in Figure 7C and D.

Figure 6. Stacked bar chart depicting a four-grade scale in which respondents subjectively rated each organ at risk. The numbers within the bars indicate the observed frequency for each of the four categories: acceptable (no modification is required) with green, minor changes are required (50% or more time savings can be expected) with yellow, major changes are required (0-50% time savings can be expected) with orange, and not acceptable (faster to delineate from scratch) with red.

Figure 7. Example contours of the submandibular gland, thyroid, and esophagus on the axial computed tomography scan that required modification from scratch. (A) Submandibular gland, (B) Thyroid, (C) Thyroid and esophagus, (D) Esophagus. The contours are color-coded as follows: navy blue (left submandibular gland), cyan (right submandibular gland), ivory (thyroid), and red (esophagus).

Discussion

This study developed an atlas model of the ABAS and reduced the work time using automated operations in automatic delineation software by scripting. Further, delineation accuracy was evaluated both geometrically and subjectively. Considering the DSC, HD value, and subjective four-grade scale, the delineation accuracy of some OARs required improvement whereas others were satisfactory.

This study aimed to evaluated not only the time reduction to delineate OARs, but also to perform the operations within the software using scripting function. It had already been reported that only auto contouring had reduced delineation time. For example, Gibbons et al. examined the influence of auto-segmentation on a clinical radiation therapy workflow, with the same indicators as in this study, using atlas auto segmentation, and reported a 42% reduction from 6.8 min to 3.9 min (17). Compared to this report, this study evaluated accuracy in more head and neck OARs and was able to reduce the overall time by 49%, including operational time (17). Thus, it is assumed that automation using scripting is able to enhance time reduction. It was shown that the time efficiency improves if the accuracy of automatically delineated contours is high (17). The contours delineated by the model created in this study have poor accuracy for some OARs, such as the chiasm, retina, lens, optic nerve, submandibular gland, thyroid, and esophagus. Further time reduction can be expected if the accuracy for these OARs is improved.

In small OARs, such as the chiasm, retina, lens, and optic nerve, the DSCs were lower than those of other OARs. Urago et al. reported that the DSCs of the chiasm, optic nerve, and retina were approximately 0.4, 0.8, and 0.8, respectively; the atlas model created in their study was considered less accurate (18). The HD value was short and delineated near the OARs because these OARs were small, with only a few slices, and the degree of correction was not significantly poor. Thus, the time required to modify the contours was considered small. The quality of CT images affects the contour delineation performance. Hence, the performance of automatic delineation algorithms may also depend on the quality of CT images. Huang et al. used an atlas model with a slice thickness of 2.5 or 3 mm to investigate the accuracy of CT images with various slice thicknesses (19). The DSC of CT images with a slice thickness larger than that of the model improved when the slice thickness was reduced; however, that of CT images with slice thicknesses smaller than 3 mm did not improve, even when the slice thickness was reduced. The accuracy of a model with a narrower slice thickness has not been reported and is therefore a topic for future research. However, for treatment planning CT in HNC, a slice thickness of 3 mm or less is recommended (20,21); this may not be necessary owing to the small modification effort.

For the submandibular gland, thyroid, and esophagus, the DSC and HD values were approximately 0.7 and less than 1 cm, respectively, indicating a satisfactory geometric evaluation. As shown in Figure 7, their contours were different from the manual contours, and all slices required modifications on every axial slice. It was deemed faster to modify them from scratch. ABAS is based on rigid and deformable registration (22); therefore, it is affected by differences in the positions and shapes of organs and changes in image contrast (23,24). Therefore, the shape and influence of contrast media on CT values may affect the deformable image registration (DIR) accuracy. Furthermore, these OARs vary widely in shape between individuals, and the esophagus may also be affected by its curvature and air. In addition, the cases enrolled in the model (approximately 10-30) represent only a limited range of anatomical changes (25,26). Therefore, these OARs could not be delineated in the cases enrolled in the model because a limited range of models could represent typical contours. The DIR between pre- and post-contrast images decreases the Dice coefficient compared with the DIR between non-contrast images (23,27). The CT images used in this study had CT values of the OARs and the vessels around these OARs, which may have affected the accuracy of the atlas segmentation.

Conclusion

The ABAS reduced the time required to delineate OAR contours in HNC. In addition, scripted automation makes it more efficient to operate within the software and reduces workload and delineation time. Further efficiency can be expected through improved delineation accuracy.

Funding

No funding was received for conducting this study.

Conflicts of Interest

The Authors declare that there are no conflicts of interest in relation to this study.

Authors’ Contributions

Yukari N. and Shingo O. designed the study, the main conceptual ideas, and the proof outline. Yukari N., Shingo O., Toshiki I., Akira M., Naoyuki K., Takahisa N., Kazunori T., and Yutaro Y. collected the data. Yukari N. wrote the manuscript with support from Shingo O., Toshiki I., and Naoyuki K. All Authors discussed the results and commented on the manuscript.

Acknowledgements

The Authors would like to thank the oncologists and radiotherapists of Osaka International Cancer Institute, Osaka, Japan for the technological assistance of contour delineation.