Quantitative and Qualitative Review of Auto-Segmentations

Performance on Challenge Test Set

The model’s performance was evaluated on the live test challenge data by comparing the ground truth and auto-segmented volumes using the DSC, MSD, and 95HD. These results are summarized in Table 12.3 and are comparable to other published results using this dataset [3, 10]. Overall, the auto-segmented volumes agreed well with the ground truth volumes. The most challenging volume to auto-segment was the esophagus. This was expected as the esophagus is a difficult volume to segment manually due to the elasticity of this volume and the potential lack of tissue contrast which prevents a clear visualization of this organ’s boundaries.

For the lungs, most disagreement between the ground truth and auto-segmented volumes was found to be within the most caudal and cranial extent of the lungs. These differences were minor for most cases (Figure 12.5); however, some disagreement was found for cases (LCTSC-S2-203 and LCTSC-S3-203, Figure 12.6) where lung volumes were affected by patients’ disease. When considering the heart, the worst performing auto-segmentation (based on DSC, DSC = 0.88) was on case LCTSC-S3-204 (Figure 12.7). Here, the auto-segmentation lacks anterior coverage of the pericardium (top left and bottom left panels) which is properly covered by the ground truth heart

Illustration of К-means clustering approach to generate cropped CT image patches for testing

FIGURE 12.4 Illustration of К-means clustering approach to generate cropped CT image patches for testing. The left and right panels represent the approach in 3D and 2D (single slice for illustration purposes), respectively. Using the coarse segmentation masks, clusters were generated, and the cluster centers (dark circles in 2D illustration) were used to center the image crop input for the fine-detail segmentation network.

TABLE 12.3

Summary of Overlap and Distance Metrics Calculated between Auto- Segmentations and Ground Truth Volumes from the Live Test Challenge Data

Lung_L

DSC

MSD (mm)

95HD (mm)

0.97

+/-

0.03

0.6

+/-

0.4

2.5

+/-

1.4

Lung_R

0.97

+/-

0.02

0.8

+/-

0.6

3.0

+/-

2.5

Heart

0.93

+/-

0.02

2.0

+/-

0.6

5.8

+/-

1.2

Esophagus

0.67

+/-

0.11

2.4

+/-

1.8

9.5

+/-

8.5

Spinal Cord

0.84

+/-

0.05

0.8

+/-

0.3

2.4

+/-

0.7

Comparison between the ground truth

FIGURE 12.5 Comparison between the ground truth (darker) and auto-segmented (lighter) lungs. The top row shows minor differences between these volumes towards the cranial extent of the lungs, whereas the bottom row highlights differences in the most caudal extent of the lungs.

volume. In some cases, there were slight differences between the ground truth and auto-segmented volumes around the apex of the heart.

When considering the spinal cord and esophagus, poorer performance (in terms of the overlap and distance metrics) was noticed compared to that found for the heart and lung segmentations. The Radiation Therapy Oncology Group (RTOG) contouring guideline, RTOG 1106, defines the spinal cord by the bony limits of the spinal canal starting at the level just below the cricoid. The auto-segmentation model produced contours that followed the spinal cord and, in some slices, failed to fully encompass the spinal canal (Figure 12.8). Interestingly, the cranial extent of auto-segmented spinal cord volumes varied widely sometimes reaching the base of the skull as shown on Figure 12.8a. After review of esophageal auto-segmentations, it was noticed that auto-segmentations often fail on the CT slices where the heart is present. As the esophagus runs behind the heart, it tends to vary widely in appearance as it is compressed between the heart vessels, vertebral bodies, and the trachea. Interestingly, the esophagus auto-segmentations were more consistent with ground truth volumes when air bubbles are present throughout the esophageal cavity (Figure 12.9).

Different Anatomical Sites

In this section, the model’s ability to auto-segment all five OARs on CT scans from non-thoracic primary tumor sites was qualitatively evaluated. A subjective evaluation of the auto-segmentation model could highlight potential problems when translating the trained thoracic OAR auto-segmen- tation model to auto-segment these normal tissues on simulation CT scans from other treatment sites.

Flighlight of disagreement between ground truth (darker) and auto-segmented (lighter) lung volumes due to the presence of different pathologies

FIGURE 12.6 Flighlight of disagreement between ground truth (darker) and auto-segmented (lighter) lung volumes due to the presence of different pathologies.

Head and Neck Scans

Head and neck radiotherapy treatment planning requires manual contours for several organs at risk and target volumes. For the past 20 years, several approaches have been proposed for the autosegmentation of these volumes [5, 11-16]. More recently, several publications using deep learning approaches have shown that head and neck normal tissue auto-segmentations can be of high quality and may not require manual edits leading to an improvement over previously proposed techniques.

From the five organs at risk available in the challenge dataset, the heart is usually not included in the FOV of head and neck simulation scans, but the spinal cord, esophagus, and both lungs are typically (at least partially) present within the CT scan. For this reason, the qualitative evaluation focuses on the auto-segmentations of the spinal cord, esophagus, and both lungs. Here, five head and neck cancer patients previously treated at MD Anderson Cancer Center with radiotherapy were randomly selected. DICOM files for the simulation CT scans were exported and auto-segmentations were produced using the two-stage deep learning model described in Section 12.2.

Visual inspection of the auto-segmentations showed similar results for the spinal cord as those observed in the challenge test data. For example, the cranial extent of the spinal cord auto-segmentations varied across the cases inspected. For some cases it was found that the spinal cord auto-segmentation reached the brainstem while for others the spinal cord was only auto-segmented partially through the head and neck region. Esophagus auto-segmentations covered the organ at risk appropriately for all cases; this was consistent with the findings on the challenge test that autosegmentations overlapped the ground truth volumes well in the cervical and upper thoracic sections of the esophagus.

Axial, coronal, and sagittal views of ground truth (darker) and auto-segmented (lighter) heart volumes

FIGURE 12.7 Axial, coronal, and sagittal views of ground truth (darker) and auto-segmented (lighter) heart volumes.

Illustration highlighting the differences between ground truth

FIGURE 12.8 Illustration highlighting the differences between ground truth (darker) and auto-segmented (lighter) spinal cord volumes. All panels (a-h) are from challenge data case LCTSC-Test-S3-204. Panels (a) and (h) show sagittal views across different planes to encompass the curvature of the spine. Panels (b-g) show axial views of the spinal cord contours. Panels (b) and (f) show typical discrepancies between ground truth and auto-segmented volumes.

Surprisingly, the lung auto-segmentations (Figure 12.10) failed to properly cover the lung volumes at the caudal edge of the CT scan. Upon inspection of the prediction probability maps, it was found that both U-net architectures used in the first stage and second stage of the model undercontoured the lung volumes in these cases. Interestingly, the most medial portions of the lungs were contoured properly in most cases, but the segmentations failed to cover the most lateral sides of the lungs even when these regions were sampled in the second stage of the model.

Comparison of ground truth

FIGURE 12.9 Comparison of ground truth (red) and auto-segmented (green) esophagus volumes. The two cases shown illustrate the difference between the 15th and 85th percentile DSC scores volumes. It is easy to visualize the disagreement for the case on the left half (LCTSC-Test-S2-203), the auto-segmentation fails to follow the curvature of the esophagus as it runs posteriorly to the heart. The auto-segmentation in the case displayed on the right half (LCTSC-Test-Sl-201) follows the trajectory of the esophagus with higher agreement.

Abdominal Scans

Simulation CT scans of the abdomen are like head and neck CT scans in that they generally only provide a partial view of the thoracic organs at risk. Here, abdominal scans include the most caudal extent of these organs, whereas head and neck scans include the most cranial extent of these volumes, as seen in Section 12.3.2.1. As the heart sits above the diaphragm, it is usually included in the scan’s FOV.

For this subsection, liver cancer patients who were previously treated with radiotherapy at MDAnderson Cancer Center were randomly selected. The thoracic auto-segmentation model was applied to these patients’ simulation CT scans. Upon visual inspection of the auto-segmentations (Figure 12.11), it was found that most auto-segmented volumes were appropriately defined on these cases. In contrast to the findings on head and neck scans (Section 12.3.2.1), the partial lung volumes were properly auto-segmented in the abdominal CT scans. The spinal cord auto-segmentations spanned from the most cranial to the most caudal extent of the CT scan for all cases. The esophagus auto-segmentations were of reasonable quality with some variability in coverage on the caudal edge near the gastroesophageal junction. In some cases, the full volume of the heart was present in the CT scan’s FOV; however, this was not true for all scans. When the heart was close to the edge of the FOV of the scan, slight inaccuracies in the heart auto-segmentations were noted, which can be appreciated in the top row (center) panel of Figure 12.11. This panel shows how the posterior portion of the heart is under-contoured as the heart volume gets closer to the most superior edge of the scan.

Breast Cancer Simulation Scans

Breast cancer patients’ simulation CT scans generally have a similar FOV to those found in thoracic cancer patients. A way thoracic and breast cancer patients’ radiotherapy simulation CT scans can differ is in how patients are positioned during simulation and treatment. For example, thoracic cancer patients are typically set up in the supine position with both arms above the head. Breast cancer patients’ setup can vary depending on the extent of the disease. At a minimum the ipsilateral arm is up, but sometimes both arms are placed above the head. An adjustable breast board is generally used to set up the patient in an incline (typically 5-15 degrees) to isolate the breast tissue below the clavicle. Specialized breast treatment immobilization devices such as breast compression paddles are often used throughout treatment.

Illustration of auto-segmented thoracic organs at risk on two

FIGURE 12.10 Illustration of auto-segmented thoracic organs at risk on two (a and b) head and neck cancer CT simulation scans. The auto-segmented spinal cord (red), esophagus (pink), and right (yellow) and left (blue) lungs are displayed. For each case (a and b) a coronal and four axial views are displayed illustrating how the lung auto-segmentations result in under-contouring of the lung volumes.

Auto-segmentations on breast cancer patient simulation CT scans were evaluated to determine the effect of the inclusion of breast cancer-specific immobilization devices and patient set up on organs at risk auto-segmentations. The inclusion of breast compression paddles is evaluated. Figure 12.12 shows a patient simulated for treatment of the left breast (a) who was later simulated for a boost treatment using a compression paddle (b). The addition of the breast paddle did not show any noticeable effects on the auto-segmentations across multiple cases.

Figure 12.13 shows two breast cancer patients who were treated using a similar setup during radiotherapy simulation. For these cases, a metal rod is used to mark and identify on the CT scan the superior border of the non-divergent fields used to treat these patients. Interestingly, it was noticed, for the case on the left column, that the presence of the rod influenced the left lung autosegmentation (blue) as can be appreciated in the axial and coronal views. Here, the lateral border of the lung is under-contoured towards on adjacent slices where the rod is present. This effect is not observed for all cases where this rod marker is used as can be seen for the case on the right column of Figure 12.13.

Auto-segmented organs at risk on abdominal CT scans from two liver cancer patients

FIGURE 12.11 Auto-segmented organs at risk on abdominal CT scans from two liver cancer patients (different patients separated by row). The auto-segmented spinal cord (red), esophagus (pink), heart (orange), and right (yellow) and left (blue) lungs are displayed on the coronal, sagittal, and axial views.

Left breast cancer patient with two simulation scans for (a) whole breast irradiation and (b) boost radiotherapy plans

FIGURE 12.12 Left breast cancer patient with two simulation scans for (a) whole breast irradiation and (b) boost radiotherapy plans.

Two breast cancer patients

FIGURE 12.13 Two breast cancer patients (different patients separated by column) using the same immobilization device which includes a metal rod to identify the non-divergent border of the tangent fields on the CT scan. The heart (red) and left (blue) and right (yellow') lung auto-segmentations are displayed on the axial and coronal views.

Different Clinical Presentations

In this section, the effect of anatomical changes to the thorax produced by pre-existing conditions (atelectasis and pleural effusions) and patient immobilization techniques are investigated, as well as the effect of the presence of contrast and implanted devices on the auto-segmentation of OARs from a deep learning model trained using the challenge dataset.

Atelectasis and Pleural Effusion

Atelectasis and pleural effusions are not uncommon in cancer patients, with some studies reporting 7-22% of lung cancer patients presenting with these conditions prior to the start of treatment [visual inspection of challenge training data showed that 5/36 (14%) of cases showed signs of atelectasis (LCTSC-S1-008, LCTSC-S1-011, LCTSC-S3-005, LCTSC-S3-008, and LCTSC-S3-012)]. Atelectasis is a condition in which the airways and air sacs in the lung collapse or do not expand properly. This condition can happen in the presence of a tumor or because of a pleural effusion which is a condition that affects the tissue that covers the outside of the lungs and lines the inside of the chest cavity. These occur when an infection, medical condition, or chest injury causes fluid, pus, blood, air, or other gases to build up in the pleural space. Both atelectasis and pleural effusions can greatly alter the anatomy within the thorax; therefore, the auto-segmented organs at risk in cases presenting with these conditions were qualitatively evaluated.

Pleural effusions resulting in total collapse of the lungs showed to have an impact on the quality of the auto-segmentations. The left column on Figure 12.14 shows axial and coronal views from a patient who presented with pleural effusion and atelectasis of the left lung. Auto-segmentations on this patient’s CT scan did not produce accurate results; the left lung auto-segmentation fails to fully cover the remaining healthy lung whereas the heart auto-segmentation extends outside of the pericardium and into the fluid in the pleura. The patient shown in the axial and coronal view's in

Auto-segmentation of organs at risk in the presence of atelectasis and/or pleural effusions

FIGURE 12.14 Auto-segmentation of organs at risk in the presence of atelectasis and/or pleural effusions. The heart (red) and left (blue) and right (yellow) lungs are displayed in the axial and coronal views for three lung cancer patients (columns).

the center panel of Figure 12.14 presented with atelectasis of the right upper lobe of the lung (top center panel). For this case, the auto-segmentation model was able to accurately define both lungs and the heart. The third case shown in Figure 12.14 (right panel) presented with atelectasis of the left lung due to the presence of a tumor. Upon visual inspection of this case, it was found that the auto-segmentations accurately contoured all organs at risk.

Presence of Motion Management Devices

Previously, Feng et al. [2] showed that the use of abdominal compressions at their institution led to an over-contouring of the heart in the caudal direction (Figure 12.2a). Here, replication of this failure mode is attempted by predicting on a case previously treated with abdominal compressions. Unfortunately, this immobilization technique is rarely used at MD Anderson Cancer Center and only a single case was found on the database to replicate their findings. For this single case, good agreement was found between the auto-segmented (green) and clinically contoured (red) heart volumes as shown in Figure 12.15. A difference between the model used in this chapter and the one used by Feng et al. is that the heart input size is larger in the SI direction (64 x 128 x 128 vs 32 x 160 x 192). It is possible that the greater field of view in the SI direction used in this chapter’s model

Comparison of ground truth (red) and auto-segmented (green) heart volumes on a lung cancer patient where abdominal compression immobilization was used

FIGURE 12.15 Comparison of ground truth (red) and auto-segmented (green) heart volumes on a lung cancer patient where abdominal compression immobilization was used.

allows for a broader view of the SI borders of the heart; however, this remains to be investigated and is outside of the scope of this analysis.

Use of Contrast and the Presence of Implanted Devices

Contrast-enhanced CT scans are commonly used in cancer diagnosis, staging, and post-treatment follow-up. In radiotherapy, the use of intravenous (IV) contrast during treatment simulation CT scan acquisition can allow for enhanced visualization of target volumes and adjacent organs at risk facilitating more accurate delineations of these volumes [17]. For non-small cell and small cell lung cancers, the National Comprehensive Cancer Network recommends the use of IV contrast with or without oral contrast for better target and organ delineation in patients with central tumors or nodal disease [18]. The presence of contrast in CT imaging changes the local distribution of HU values which can have an impact on auto-segmentations. Previous works have addressed this challenge by using training sets that contain contrast-enhanced and non-contrast CT images [19]. In this subsection, the model trained on the challenge data was used to evaluate the quality of auto-segmentations in cases where there are high density regions such as those in contrast-enhanced CT images and implanted devices.

Figure 12.16 shows three patients’ CT scans (axial and coronal or sagittal views) that illustrate the possible failure modes observed when auto-segmenting OARs in cases where contrast or hardware are present. The patient shown on the left panel of Figure 12.16 had contrast in their bowels. Here, the left lung (blue) is not auto-segmented when it is near the high-density regions in the scan (as seen on coronal and axial views). In contrast, the right lung (yellow) was accurately auto- segmented for this case. For this particular case, it was found that most CT image patches used to predict the left lung included regions where contrast was present; this was not the case for the right lung where the liver was found in the most caudal regions of a large number of image patches. The patient shown in the center panels of Figure 12.16 had a previously implanted pacemaker (as shown on the top two center panels).

CT images from a patient with contrast in the abdomen

FIGURE 12.16 CT images from a patient with contrast in the abdomen (left), pacemaker (center), and contrast in the heart (right). The heart (red) and left (blue) and right (yellow) lungs are displayed to illustrate failures in auto-segmentations.

The pacemaker leads appear as bright wires that transverse from the collar bone to the chambers of the heart. For this patient, the presence of the high-density leads traveling through the heart affected the quality of the heart auto-segmentation (as seen on the bottom axial and sagittal views). Lastly, the right panel of Figure 12.16 shows a patient who had a chest CT with contrast for radiotherapy simulation. Flcre the presence of contrast through the heart caused the deep learning- based auto-segmentation algorithm to significantly under-contour the heart volume. Interestingly, the presence of contrast did not affect the quality of the auto-segmentations of the lungs. These findings merit additional investigations to better understand the mechanisms for why the presence of high-density regions affects some auto-segmentations.

Adapting to the Unseen

When a deep learning auto-segmentation model fails to produce high-quality segmentations due to limitations on the training data (low number of training cases or limited anatomical diversity), there are several options to overcome these challenges. Feng et al. [2] showed how they adapted their previously trained convolutional neural network by retraining their model with both challenge and additional institutional data. In their study, the authors illustrate how both transfer learning and end-to-end training can improve the quality of the segmentations (Figure 12.2b) by learning patterns from the additional training data. Other plausible options include changing the architecture (see Chapters 7, 8, and 9), changing model parameters such as input size and loss function (see Chapter 10), or to change how training data is input through a model during training (see Chapter 11). All of these could have a significant impact on the resulting auto-segmentations. Lastly, curation of additional training data could help address many of the limitations observed due to diversity in anatomical presentations and/or model applications.

Discussion and Conclusions

In this chapter, a two-stage deep learning auto-segmentation model was trained using the 2017 AAPM Thoracic Auto-segmentation Challenge data and the ability of this model to produce accurate auto-segmentations on a variety of clinical uses and scenarios was evaluated. The model used a generic 3D U-net architecture that first localizes (stage 1) the organs at risk and then auto-segments (stage 2) individual organs using cropped CT scan volumes about each organ. The resulting model was validated using the challenge test dataset producing similar results to those reported in the literature.

To identify possible scenarios where the trained auto-segmentation model could fail, the model’s performance was evaluated on two types of scenarios. First, it was investigated whether using simulation CT scans from other treatment sites would have an influence on the accuracy of the autosegmentations. Potential issues were identified with auto-segmentations that are near the edge of the scan’s FOV. Several reasons could be attributed to this observation and there is a high likelihood that these issues could be resolved by using a simple 2D architecture, yet this remains to be evaluated. Another failure mode in the auto-segmentations was noticed when streaking artifacts from a breast cancer radiotherapy immobilization device were located near the lungs. Under-contouring in these regions near the artifact varied from patient to patient (Figure 12.13).

The second type of scenario investigated was the potential effect of anatomical changes to the thorax produced by pre-existing conditions (atelectasis and pleural effusions) and patient immobilization technique, as well as the effect of the presence of contrast and implanted devices on the auto-segmentation. When considering atelectasis and pleural effusions, it was found that the accuracy of the auto-segmentations was greatly reduced when there was significant buildup of fluid in the pleura. Only a single case was located at MD Anderson Cancer Center where the abdominal compression immobilization technique was used during radiotherapy simulation; all organs at risk were accurately auto-segmented for this case but additional cases are needed to confirm that slight parameter differences in approach resolved the auto-segmentation errors previously reported by Feng et al. [2]. Lastly, the presence of contrast and high-density materials from medical implants was shown to have a negative impact on the quality of the auto-segmentations.

The case study presented has a few limitations. First, a single architecture is considered highlighting issues that might not be observed with other similar 3D approaches. Second, only a limited range of possible scenarios were considered in the qualitative evaluation of this case study. Potential scenarios such as anatomical changes from surgical procedures (i.e. pulmonary lobectomy) could have significant impact on the accuracy of auto-segmentations. Lastly, the analysis presented is limited as it only uses the 2017 AAPM Thoracic Auto-segmentation Challenge data for the case study. Expanding this analysis to other publicly available datasets could confirm the shortcomings observed in this study.

In conclusion, while the model presented produced high quality auto-segmentations when compared to the challenge test data, some shortcomings were observed when applying such a model to a diverse set of scenarios and clinical uses, highlighting the limitations of the proposed model and the use of limited training datasets. Identifying failure modes of a deep learning auto-segmentation model can be a critical step in the clinical deployment of such models and could improve patient safety throughout the radiotherapy treatment planning process.

Acknowledgments

The author would like to acknowledge the support of the High Performance Computing facility at the University of Texas MD Anderson Cancer Center for providing computational resources (including consulting services) that have contributed towards the training of models and generation of data used for this book chapter.

References

  • 1. Cardenas CE, Yang J, Anderson BM, Court LE. Brock KB. Advances in auto-segmentation. Semin Radial Oncol 2019;29(3): 185-197. doi:10.1016/j.semradonc.2019.02.001
  • 2. Feng X, Bernard ME, Hunter T, Chen Q. Improving accuracy and robustness of deep convolutional neural network based thoracic OAR segmentation. Phys Med Biol 2020:65(7). doi:10.1088/1361-6560/ ab7877
  • 3. Feng X, Qing K, Tustison NJ, Meyer CH. Chen Q. Deep convolutional neural network for segmentation of thoracic organs-at-risk using cropped 3D images. MedPhys 2019;46(5):2169—2180. doi:10.1002/ mp. 13466
  • 4. £iyek 6, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O. 3D U-net: learning dense volumetric segmentation from sparse annotation. Led Notes Comput Sci 2016;9901 LNCS:424-432. doi:10.1007/978-3-319-46723-8_49
  • 5. Rhee DJ, Cardenas CE, Elhalawani H, et al. Automatic detection of contouring errors using convolutional neural networks. MedPhys September 2019:mp.l3814. doi:10.1002/mp,13814
  • 6. Balagopal A, Kazemifar S, Nguyen D, et al. Fully automated organ segmentation in male pelvic CT images. Phys Med Biol 2018:63(24). doi:10.1088/1361-6560/aafl lc
  • 7. Wang Y, Zhao L, Wang M, Song Z. Organ at risk segmentation in head and neck CT images using a two-stage segmentation framework based on 3D U-Net. IEEE Access. 2019;XX:1—1. doi: 10.1109/ access.2019.2944958
  • 8. Karimi D, Samei G, Kesch C, Nir G, Salcudean SE. Prostate segmentation in MRI using a convolutional neural network architecture and training strategy based on statistical shape models. Ini J Comput Assist Radiol Surg 2018;13(8):1211—1219. doi:10.1007/sl 1548-018-1785-8
  • 9. Yang Y, Jiang H, Sun Q. A multiorgan segmentation model for CT volumes via full convolution-deconvolution network. Biomed Res Int 2017:2017:1-9. doi: 10.1155/2017/6941306
  • 10. Um H, Jiang J, Thor M, et al. Multiple resolution residual network for automatic thoracic organs-at-risk segmentation from CT. 2020:1-5. http://arxiv.org/abs/2005.13690.
  • 11. Yang J, Zhang Y, Zhang L, Dong L. Automatic segmentation of parotids from CT scans using multiple atlases. Med Image Anal Clin A Gd Chall. CreateSpace Independent Publishing Platform, 2010:323- 330. www.amazon.com/Medical-Image-Analysis-Clinic-Challenge/dp/1453759395
  • 12. Stapleford LJ, Lawson JD, Perkins C, et al. Evaluation of automatic atlas-based lymph node segmentation for head-and-neck cancer. Int J Radiat Oncol Biol Phys 2010;77(3):959—966. doi: 10.1016/j. ijrobp.2009.09.023
  • 13. Cardenas CE, McCarroll RE, Court LE, et al. Deep learning algorithm for auto-delineation of high-risk oropharyngeal clinical target volumes with built-in dice similarity coefficient parameter optimization function. Int J Radiat Oncol Biol Phys 2018;101(2):468—478. doi: 10.1016/j.ijrobp.2018.01.114
  • 14. McCarroll R, Yang J, Cardenas CE, et al. Machine learning for the prediction of physician edits to clinical auto-contours in the head-and-neck. Med Phys 2017;44(6):3160.
  • 15. Cardenas CE, Anderson BM, Aristophanous M, et al. Auto-delineation of oropharyngeal clinical target volumes using 3D convolutional neural networks. Phys Med Biol 2018:63(21). doi:10.1088/1361-6560/ aae8a9
  • 16. Men K, Chen X, Zhang Y, et al. Deep deconvolutional neural network for target segmentation of nasopharyngeal cancer in planning computed tomography images. Front Oncol 2017;7(December):l-9. doi: 10.3389/fonc.2017.00315
  • 17. Bae KT. Intravenous contrast medium administration and scan timing at CT: considerations and approaches. Radiology 2010;256(1):32—61. doi: 10.1148/radiol. 10090908
  • 18. National Comprehensive Cancer Network. Clinical Practice Guidelines in Oncology for Non-Small Cell Lung Cancer. 2020. www.nccn.org/professionals/physician_gls/pdf/nscl.pdf.
  • 19. Anderson BM, Lin EY, Cardenas C, et al. Automated contouring of variable contrast CT liver images. Adv Radiat Oncol 2020. doi:10.1016/j.adro.2020.04.023
 
Source
< Prev   CONTENTS   Source   Next >