https://orcid.org/0000-0003-2027-7472
Figures
Abstract
Background
Total marrow irradiation (TMI) and total marrow and lymphoid irradiation (TMLI) have the advantages. However, delineating target lesions according to TMI and TMLI plans is labor-intensive and time-consuming. In addition, although the delineation of target lesions between TMI and TMLI differs, the clinical distinction is not clear, and the lymph node (LN) area coverage during TMI remains uncertain. Accordingly, this study calculates the LN area coverage according to the TMI plan. Further, a deep learning-based model for delineating LN areas is trained and evaluated.
Methods
Whole-body regional LN areas were manually contoured in patients treated according to a TMI plan. The dose coverage of the delineated LN areas in the TMI plan was estimated. To train the deep learning model for automatic segmentation, additional whole-body computed tomography data were obtained from other patients. The patients and data were divided into training/validation and test groups and models were developed using the “nnU-NET” framework. The trained models were evaluated using Dice similarity coefficient (DSC), precision, recall, and Hausdorff distance 95 (HD95). The time required to contour and trim predicted results manually using the deep learning model was measured and compared.
Results
The dose coverage for LN areas by TMI plan had V100% (the percentage of volume receiving 100% of the prescribed dose), V95%, and V90% median values of 46.0%, 62.1%, and 73.5%, respectively. The lowest V100% values were identified in the inguinal (14.7%), external iliac (21.8%), and para-aortic (42.8%) LNs. The median values of DSC, precision, recall, and HD95 of the trained model were 0.79, 0.83, 0.76, and 2.63, respectively. The time for manual contouring and simply modified predicted contouring were statistically significantly different.
Citation: Choi HS, Kang H-C, Chie EK, Shin KH, Chang JH, Jang B-S (2024) Assessment of lymph node area coverage with total marrow irradiation and implementation of total marrow and lymphoid irradiation using automated deep learning-based segmentation. PLoS ONE 19(3): e0299448. https://doi.org/10.1371/journal.pone.0299448
Editor: Minsoo Chun, Chung-Ang University Gwangmyeong Hospital, REPUBLIC OF KOREA
Received: November 28, 2023; Accepted: February 10, 2024; Published: March 8, 2024
Copyright: © 2024 Choi et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The models developed based on “nnU-NET” in this study are available in the Zenodo repository (doi.org/10.5281/zenodo.7839889).
Funding: This research was supported by a grant from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health and Welfare, Republic of Korea (#HI20C2092) to Eui Kyu Chie and the National R&D Program for Cancer Control through the National Cancer Center (NCC) funded by the Ministry of Health & Welfare, Republic of Korea (HA22C0044) to Kyung Hwan Shin. The study sponsors had no study involvement. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Abbreviations: BMI, body mass index; CT, Computed tomography; DSC, Dice similarity coefficient; HD95, Hausdorff distance 95; IQR, interquartile range; LN, Lymph node; OAR, Organs at risk; RT, Radiotherapy; TBI, Total body irradiation; TMI—TMI, Total marrow irradiation; TMLI, Total marrow and lymphoid irradiation
Introduction
Total marrow irradiation (TMI) and total marrow lymphoid irradiation (TMLI) are both treatments for conditioning regimens prior to allogeneic and autologous hematopoietic stem cell transplantation for treating blood cancers (e.g., leukemia, lymphoma, and myeloma [1]). Total body irradiation (TBI) is a conventional treatment for the same purpose [2]; however, it can cause multiple toxicities despite limited prescribed radiation doses. Although it affords certain advantages over a chemotherapy monotherapy regimen [3], TBI usage has declined [4] due to its toxicity and the necessity for specialized facilities. However, advances in radiotherapy (RT) techniques have reduced the exposure of organs at risk (OAR) to radiation [5]. For example, in some clinical studies, TMLI has shown promising results, such as lower incidence of radiation-related toxicity and extramedullary relapse [6]. This demonstrates the potential for increasing prescription doses to target lesions [7].
Both TMI and TMLI can be used as pretreatments; however, a major difference is observed between them. The former only targets the bone marrow, whereas the latter targets both the bone marrow and major lymph nodes (LNs). Despite these different target lesions, the clinical variation between TMI and TMLI remains unknown; further research is required to compare these therapies. Moreover, planning and conducting clinical studies for TMI and TMLI is challenging due to their low prevalence and the lengthy follow-up required to demonstrate significant differences. Therefore, in such cases, it may be cautiously posited that clinical differences can be indirectly inferred through differences in dose distribution.
And if clinical differences exist, the delineation of the whole-body bone marrow, LN areas, and OARs is essential for TMLI treatment planning. However, this planning process is time-consuming. One study reports that approximately 12–16 h per patient are necessary [8]. Hence, the process is labor-intensive and poses a challenge for clinical adoption Recent advancements in RT have increasingly utilized deep learning technologies. These include applications in generating planning CT data from diagnostic MRIs [9], assisting in target contouring [10], creating treatment plans [11], and predicting radiotherapy outcomes [12]. A noteworthy aspect of this evolution is the significant time reduction in target contouring achieved through deep learning, as confirmed in a study [10]. This progress suggests a potential decrease in the time required for contouring in TMLI. The “nnU-NET” framework, prominent among deep learning models in these studies, stands out for its excellent performance and user-friendliness, making it accessible even to those without expert-level knowledge in computer programming [13].
In this study, dose coverage is estimated by delineating whole-body LN areas based on the computed tomography (CT) data of patients treated according to the TMI plan. Furthermore, a deep learning model is trained and evaluated for auto-contouring whole-body LN areas using a deep learning framework named “nnU-NET”.
Materials and methods
Patients and target structure delineation
In this study, the CT data from patients treated with TMI in 2017 and the CT data of patients who underwent whole-body CT scans from January 1, 2021 to December 31, 2021 at our institution were collected. The median age of the patients at the time of TMI was 40 years (interquartile range (IQR): 28–54 years), and 85% were male; 69% were diagnosed with leukemia. The median height, weight, and body mass index (BMI) of the patients who of the TMI patients were 170.8 cm (IQR: 164.8–175.8 cm), 69.1 kg (IQR: 63.5–80.3 kg), and 23.9 kg/cm2 (IQR: 23.3–25.7 kg/cm2), respectively. Regarding the prescribed doses for the 13 patients, one received 8 Gy in 4 fractions, another received 12 Gy in 4 fractions, and the remaining 11 patients received 10 Gy in 5 fractions.
To train the deep learning model, data were collected from 13 patients who underwent whole-body CT at our institution between January 1, 2021, and December 31, 2021. The median age of patients at the time of the CT scan was 64 years (IQR: 56–72 years, with 46% being male, and 69% undergoing whole-body CT for multiple myeloma diagnosis. Their median height, weight, and BMI were 162.25 cm (IQR: 150.75–164.5 cm), 61 kg (IQR: 53.5–68.5 kg), and 25.1 kg/cm2 (IQR: 23.4–25.7 kg/cm2), respectively. All collected CT data were non-contrast enhanced CT scans and the average pixel size was 1.113mm x 1.113mm, and the slice thickness was consistently 5mm.
Eighteen LN areas were manually delineated on each patient’s CT data by a radiation oncologist and then confirmed by two radiation oncologists. These LN areas included the cervical (Lt/Rt), axillary (Lt/Rt), and mediastinal LNs (Sup. /Inf.); para-aortic LN area; common iliac LN area (Lt/Rt); external iliac LN area (Lt/Rt); internal iliac LN area (Lt/Rt); obturator LN area (Lt/Rt); presacral LN area; and inguinal LN area (Lt/Rt). To delineate each LN area, delineation guidelines for head and neck, breast, lung, and cervical cancers were referenced [14–19].
Dose estimation for LN areas treated by TMI
To calculate the dose received by each LN area using the TMI plan, information from patients who underwent TMI was used. Dose planning from previous TMI treatments was reloaded, and a manually delineated LN area with margins(5mm) for planning target volume was used in each plan to calculate the dose for each area. For each patient, Vx%, which is the volume of the organ receiving x% or more of the prescription dose, was calculated.
Development of deep learning model
To train the deep learning model for medical image segmentation tasks, the “nnU-Net” framework is utilized [13]. This framework is characterized by its ability to function without extensive data pre-processing or parameter tuning. This is achieved through an “out-of-the-box” approach that uses a pretrained network and requires only minimal fine-tuning of target data. Hence, the framework is an ideal choice for users who lack expertise in image segmentation and deep learning. The “nnU-Net” framework has been extensively evaluated in a variety of imaging tasks and shown to achieve state-of-the-art performance especially in medical imaging tasks. The patients from whom CT data were collected were divided, with 16 individuals allocated to the training group, 4 to the validation group, and 6 to the test group. Training was performed by fivefold cross-validation using a 4:1 split of CT data from the training/validation group; the model with the best evaluation parameter value was selected. Training and testing were performed using a workstation equipped with an NVIDIA RTX A6000 with 48GB of graphics processing unit memory.
The models developed based on “nnU-NET” in this study are available in the Zenodo repository (doi.org/10.5281/zenodo.7839889).
Model evaluation
The trained models were evaluated using dice similarity coefficient (DSC), precision, recall, and Hausdorff distance 95 (HD95). The DSC is used to quantify the overlap between an observer and consensus contours; it is commonly employed as an index for image segmentation evaluation. The coefficient ranges from 0 (indicating no similarity) to 1 (indicating a perfect match). When the volume of ground truth is “G” and the volume of prediction by a trained deep learning model is “P,” the DSC, precision, and recall are calculated as 2×|G∩P|/(|G|+|P|), |G∩P|/|G|, and |G∩P|/|P|, respectively. To measure the distance between two sets, HD95 calculates the farthest point between the two. The HD95 approach is used because it is less sensitive to outliers. It measures the distance between two sets in which one set contains 95% or more of the points found in the other set. The evaluation metrics were calculated by aggregating values from LN areas, including both left/right and superior/inferior divisions. For the patients in the test group, the times necessary for manually delineating whole-body LN areas and accurately trimming the predicted LN areas are measured and then compared. This process is presented in Fig 1.
The CT data of 26 patients were used. The patients were divided into 20 and 6 as training/validation and test groups, respectively. Abbreviations—TMI: total marrow irradiation; CT: computer tomography.
Ethics statement
This study was approved by the Institutional Review Board of Seoul National University Hospital (IRB No. H-2204-050-1314). Personally identifiable information required for the CT data collection process was anonymized after collection. This study was approved by the IRB committee to waive consent because it is a retrospective medical record review study, obtaining participants’ consent is not practical at the time of the study, and the risk to participants of waiving consent is extremely low.
Results
Dose estimation for LN areas treated by TMI
Table 1 and Fig 2 summarizes the dose distribution in the manually delineated LN using the TMI plan. The median values of V100%, V95%, and V90% of the entire manually contoured LN area were 46.5% (IQR: 33.0–58.7%), 62.4% (IQR: 49.5–78.3%), and 75.3% (IQR: 63.8–89.6%), respectively. The presacral, axilla, and obturator LN areas had relatively high V100% median values of 71.1% (IQR: 62.2–80.8%), 61.2% (IQR: 57.2–66.5%), and 60.3% (IQR: 53.3–84.4%), respectively. In contrast, the inguinal, external iliac, and para-aortic LN areas had relatively low V100% median values of 14.7% (IQR: 7.5–20.7%), 21.8% (IQR: 15.8–46.3%), and 42.8% (IQR: 37.8–46.5%), respectively. Similar trends were observed for V95% and V90%.
Box and whisker plot representation of the Vx% for each lymph node area. Abbreviations—Vx%: volume receiving x% or more of the prescription dose.
Development of deep learning model
Out of a total of 26 patient datasets, 20 patients were assigned to the training/validation group. Within this group, the 20 patient datasets were divided in a 4:1 ratio to facilitate 5-fold cross-validation. The model that exhibited the highest performance during this cross-validation process was then selected. Following this selection, the best-performing model was further evaluated using the test group. Fig 3 shows an example of manually delineated LN areas versus the LN areas predicted by the deep learning model in a patient. The blue and red lines denote manually delineated areas and areas predicted using the deep learning model, respectively. Table 2 shows the mean volumes of manually delineated LN areas and predicted LN areas for patients in the test group by cc.
(A) Manually delineated major lymph node areas of entire body of one patient with labels and (B) delineated lymph node area in one patient (blue lines denote manually contoured LNs, and red lines represent predictions obtained using deep learning): (1) cervical LN area; (2) axillary and mediastinal LN areas; (3) para-aortic LN area; and (4) iliac and presacral LN areas; (5) inguinal LN areas. Abbreviation—LN: Lymph node.
Model evaluation
Similarity scores between the manually delineated and predicted LN areas were estimated. As shown in Table 3 and Fig 4, the median values of DSC, precision, recall, and HD95 for the entire LN area are 0.79 (IQR: 0.70–0.84), 0.83 (IQR: 0.75–0.89), 0.76 (IQR: 0.68–0.85), and 2.63 pixels (IQR: 2.0–4.58 pixels). Among them, the cervical, inguinal, and axillary LN areas had high median DSC values of 0.85 (IQR: 0.78–0.87), 0.85 (IQR: 0.82–0.88), and 0.84 (IQR: 0.76–0.88), respectively. Conversely, the mediastinal, presacral and internal iliac LN areas had low median DSC values of 0.68 (IQR: 0.64–0.79), 0.68 (IQR: 0.48–0.77), 0.74 (IQR: 0.69–0.80), respectively. The precision was highest in the inguinal LN area, while the recall was highest in the axillary LN area. HD95 ranged from 2 and 3 pixels in most LN areas, with higher values indicating worse similarity. The presacral LN region and the para-aortic LN region had poor evaluation values (12.77 pixels (IQR: 1.65–22.32 pixels) and 5.04 pixels (IQR: 3.75–20.19 pixels), respectively).
(A) Dice similarity coefficient, precision, and recall of evaluation results; (B) Hausdorff 95 (HD95) values from evaluation results (unit: pixel). Abbreviations—DSC: Dice similarity coefficient; HD95: Hausdorff distance 95%.
Significant time differences between delineating the whole-body LN areas and trimming the predicted LN areas in the six test groups were observed (p < 0.001); the median times were 70′ 20″ (IQR: 64′ 54″–78′ 27″) and 30′ 16″ (IQR: 27′ 3″–32′ 37″), respectively.
Discussion
This study aims to calculate the dose coverage of the LN area when treated according to the TMI plan, which has not been previously implemented. Results show that the inguinal, external iliac, and para-aortic LN areas have lower coverage than expected. This is probably due to the anatomical location of the LNs away from the bone. Targeting whole-body bones can be an appropriate pre-transplant conditioning regimen for hematopoietic stem cell transplantation given that most hematopoietic stem cells are present in the bone marrow. However, because malignant cells are also located in major LNs, additional targeting of an LN area may provide better clinical outcomes. Nevertheless, comparative studies have shown that TMI and TMLI have similar clinical outcomes as TBI [6]; however, but no prospective or retrospective studies comparing the clinical outcomes of TMI and TMLI have been conducted. Future studies are necessary to compare the TMI and TMLI plans in terms of the variable dose parameters of OARs and dose coverage.
The second objective of the current study is to develop a deep learning model for the automatic delineation of whole-body LN areas because manual delineation is a time-consuming and labor-intensive process when TMLI is delivered. A model is devised and subsequently trained and tested using the data from the 26 patients treated at the institution based on the “nnU-NET” deep learning framework. Although the amount of training data was small, the model performance and quality of the contours were acceptable for TMI planning. The predicted presacral and para-aortic LN areas are observed to have higher HD95 values than the other LN areas. This may be because these two LN areas have no laterality; consequently, the effect of training using few samples on these areas is less significant than that on LN areas with laterality. Nevertheless, the automatic delineation of whole-body LN areas using the developed model was demonstrated to be significantly more time-effective than manual delineation and had the potential to reduce the time required for total treatment planning.
Several papers on auto-contouring models using deep learning have been published, such as a model that auto-contours the target lesions of the head and neck, thorax, rectum, cervix, prostate, and heart structures [10,20–25]. In TMLI, target areas include bone and LN areas. For bone, due to the significant density difference with surrounding tissues, generally favorable results are observed [25]. Conversely, in LN areas, the density contrast with adjacent tissues is not as pronounced, leading to varied outcomes in different studies. Among these, Rhee et al. [22]. employed deep learning to automatically segment the Clinical Target Volume (CTV) in cervical cancer, achieving a mean DSC and HD95 of 0.81 and 2.09, respectively, for nodal CTV, and 0.76 and 2.00 for PAN CTV. Similarly, Cardenas et al. [21]. Developed an auto-segmentation model for LN CTV in head and neck cancer patients, yielding a mean DSC between 0.843 and 0.909. Our study demonstrated a median DSC of 0.79 and a median HD95 of 2.63, presenting a competitive performance in comparison with existing studies.
In contrast to these prevalent region-specific studies, whole-body auto-contouring models have received less attention in the current research landscape. Chen et al. [25]. Reported a model that can auto-contour whole-body organs. Most models utilize a limited range of CT data for each organ, or the entire body is divided into regions and trained separately to create each model. In contrast, a model based on whole-body CT data is developed in the current study; thus, a prediction for a body LN area can be immediately provided.
One study has shown that increasing the radiation dose in TBI results in better disease control [26]. Although disease control was improved, it was not linked to improved overall survival. This might be due to the toxicity of high doses of OARs. Thus, TMI and TMLI have been proposed as alternatives to reduce toxicity [5] despite the absence of published guidelines. Recently, Dei et al. reported the variability in TMLI target delineation [27]. The study found that inter-observer and intra-observer variabilities were reduced when guidelines for delineating each regional LN area were provided. Dei et al. provided documented guidelines for the current practice. Additionally, if a deep learning model can be used to provide patient-specific LN area delineation consistently, reduced variability in target delineation may be expected in TMLI planning.
One of the limitations of the current study is the small number of patients. In addition to the data of patients treated with TMI at our institution, whole-body CT data have been used as additional training data. This implies that the patient posture and image resolution between the RT and CT plans differ. Thus, further studies in the future are required for external validation using federated learning with other institutions. Another limitation of this study is the lack of international standards for TMLI target delineation. Even when the LNs are delineated according to the available contouring guidelines for several organs, concerns regarding the LN areas that may be targeted for TMLI persist. Additional fine-tuning of the model is necessary upon the availability of relevant guidelines.
Conclusions
The dose coverage in the inguinal, external iliac, and para-aortic LN areas is suboptimal if a TMI treatment plan is implemented. The current research demonstrates that the automatic delineation of LN areas using deep learning can facilitate the implementation of the TMLI plan.
References
- 1. Wong JYC, Filippi AR, Scorsetti M, Hui S, Muren LP, Mancosu P. Total marrow and total lymphoid irradiation in bone marrow transplantation for acute leukaemia. Lancet Oncol. 2020 Oct 1;21[10]:e477–87. pmid:33002443
- 2. Copelan EA. Hematopoietic Stem-Cell Transplantation. N Engl J Med. 2006 Apr 27;354[17]:1813–26. pmid:16641398
- 3. Peters C, Dalle JH, Locatelli F, Poetschger U, Sedlacek P, Buechner J, et al. Total Body Irradiation or Chemotherapy Conditioning in Childhood ALL: A Multinational, Randomized, Noninferiority Phase III Study. J Clin Oncol. 2021 Feb;39[4]:295–307. pmid:33332189
- 4. Hong S, Barker C, Klein JP, Shaw P, Bredeson C, Angelina A, et al. Trends in Utilization of Total Body Irradiation (TBI) Prior to Hematopoietic Cell Transplantation (HCT) Worldwide. Biol Blood Marrow Transplant. 2012 Feb 1;18[2]:S336–7.
- 5. Han C, Liu A, Wong JYC. Target Coverage and Normal Organ Sparing in Dose-Escalated Total Marrow and Lymphatic Irradiation: A Single-Institution Experience. Front Oncol [Internet]. 2022 [cited 2023 Apr 17];12. Available from: https://www.frontiersin.org/articles/10.3389/fonc.2022.946725 pmid:35957914
- 6. Kong F, Liu S, Liu L, Pi Y, Pei Y, Xu D, et al. Clinical study of total bone marrow combined with total lymphatic irradiation pretreatment based on tomotherapy in hematopoietic stem cell transplantation of acute leukemia. Front Oncol [Internet]. 2022 [cited 2023 Apr 15];12. Available from: https://www.frontiersin.org/articles/10.3389/fonc.2022.936985 pmid:36052229
- 7. Wong JYC, Liu A, Schultheiss T, Popplewell L, Stein A, Rosenthal J, et al. Targeted Total Marrow Irradiation Using Three-Dimensional Image-Guided Tomographic Intensity-Modulated Radiation Therapy: An Alternative to Standard Total Body Irradiation. Biol Blood Marrow Transplant. 2006 Mar;12[3]:306–15. pmid:16503500
- 8. Schultheiss TE, Wong J, Liu A, Olivera G, Somlo G. Image-guided total marrow and total lymphatic irradiation using helical tomotherapy. Int J Radiat Oncol Biol Phys. 2007 Mar 15;67[4]:1259–67. pmid:17336225
- 9. Hsu SH, Cao Y, Huang K, Feng M, Balter JM. Investigation of a method for generating synthetic CT models from MRI scans of the head and neck for radiation therapy. Phys Med Biol. 2013 Jan;58[23]:8419. pmid:24217183
- 10. Lustberg T, van Soest J, Gooding M, Peressutti D, Aljabar P, van der Stoep J, et al. Clinical evaluation of atlas and deep learning based automatic contouring for lung cancer. Radiother Oncol. 2018 Feb 1;126[2]:312–7. pmid:29208513
- 11. Brown PD, Gondi V, Pugh S, Tome WA, Wefel JS, Armstrong TS, et al. Hippocampal Avoidance During Whole-Brain Radiotherapy Plus Memantine for Patients With Brain Metastases: Phase III Trial NRG Oncology CC001. J Clin Oncol. 2020 Apr 1;38[10]:1019–29. pmid:32058845
- 12. Kang J, Schwartz R, Flickinger J, Beriwal S. Machine Learning Approaches for Predicting Radiation Therapy Outcomes: A Clinician’s Perspective. Int J Radiat Oncol. 2015 Dec 1;93[5]:1127–35. pmid:26581149
- 13. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021 Feb;18[2]:203–11. pmid:33288961
- 14. Grégoire V, Ang K, Budach W, Grau C, Hamoir M, Langendijk JA, et al. Delineation of the neck node levels for head and neck tumors: A 2013 update. DAHANCA, EORTC, HKNPCSG, NCIC CTG, NCRI, RTOG, TROG consensus guidelines. Radiother Oncol. 2014 Jan 1;110[1]:172–81. pmid:24183870
- 15. Offersen BV, Boersma LJ, Kirkove C, Hol S, Aznar MC, Sola AB, et al. ESTRO consensus guideline on target volume delineation for elective radiation therapy of early stage breast cancer, version 1.1. Radiother Oncol. 2016 Jan 1;118[1]:205–8. pmid:26791404
- 16. Lynch R, Pitson G, Ball D, Claude L, Sarrut D. Computed tomographic atlas for the new international lymph node map for lung cancer: A radiation oncologist perspective. Pract Radiat Oncol. 2013 Jan 1;3[1]:54–66. pmid:24674264
- 17. Keenan LG, Rock K, Azmi A, Salib O, Gillham C, McArdle O. An atlas to aid delineation of para-aortic lymph node region in cervical cancer: Design and validation of contouring guidelines. Radiother Oncol. 2018 Jun 1;127[3]:417–22. pmid:29523410
- 18. Taylor A, Rockall AG, Reznek RH, Powell MEB. Mapping pelvic lymph nodes: Guidelines for delineation in intensity-modulated radiotherapy. Int J Radiat Oncol. 2005 Dec 1;63[5]:1604–12. pmid:16198509
- 19. Kim CH, Olson AC, Kim H, Beriwal S. Contouring inguinal and femoral nodes; how much margin is needed around the vessels? Pract Radiat Oncol. 2012 Oct 1;2[4]:274–8. pmid:24674164
- 20. Hague C, McPartlin A, Lee LW, Hughes C, Mullan D, Beasley W, et al. An evaluation of MR based deep learning auto-contouring for planning head and neck radiotherapy. Radiother Oncol. 2021 May 1;158:112–7. pmid:33636229
- 21. Cardenas CE, Beadle BM, Garden AS, Skinner HD, Yang J, Rhee DJ, et al. Generating High-Quality Lymph Node Clinical Target Volumes for Head and Neck Cancer Radiation Therapy Using a Fully Automated Deep Learning-Based Approach. Int J Radiat Oncol. 2021 Mar 1;109[3]:801–12. pmid:33068690
- 22. Rhee DJ, Jhingran A, Rigaud B, Netherton T, Cardenas CE, Zhang L, et al. Automatic contouring system for cervical cancer using convolutional neural networks. Med Phys. 2020;47[11]:5648–58. pmid:32964477
- 23. Zabel WJ, Conway JL, Gladwish A, Skliarenko J, Didiodato G, Goorts-Matthews L, et al. Clinical Evaluation of Deep Learning and Atlas-Based Auto-Contouring of Bladder and Rectum for Prostate Radiation Therapy. Pract Radiat Oncol. 2021 Jan 1;11[1]:e80–9. pmid:32599279
- 24. Chin V, Finnegan RN, Chlap P, Otton J, Haidar A, Holloway L, et al. Validation of a Fully Automated Hybrid Deep Learning Cardiac Substructure Segmentation Tool for Contouring and Dose Evaluation in Lung Cancer Radiotherapy. Clin Oncol [Internet]. 2023 Mar 11 [cited 2023 Apr 20]; Available from: https://www.sciencedirect.com/science/article/pii/S0936655523001127 pmid:36964031
- 25. Chen W, Wang C, Zhan W, Jia Y, Ruan F, Qiu L, et al. A comparative study of auto-contouring softwares in delineation of organs at risk in lung cancer and rectal cancer. Sci Rep. 2021 Nov 26;11[1]:23002. pmid:34836989
- 26. Clift RA, Buckner CD, Appelbaum FR, Sullivan KM, Storb R, Thomas ED. Long-Term Follow-Up of a Randomized Trial of Two Irradiation Regimens for Patients Receiving Allogeneic Marrow Transplants During First Remission of Acute Myeloid Leukemia. Blood. 1998 Aug 15;92[4]:1455–6. pmid:9694737
- 27. Dei D, Lambri N, Stefanini S, Vernier V, Brioso RC, Crespi L, et al. Internal Guidelines for Reducing Lymph Node Contour Variability in Total Marrow and Lymph Node Irradiation. Cancers. 2023 Jan;15[5]:1536. pmid:36900326