Radiol Oncol 2024; 58(2): 289-299. doi: 10.2478/raon-2024-0018 289 research article Dosimetry and efficiency comparison of knowledge-based and manual planning using volumetric modulated arc therapy for craniospinal irradiation Wei-Ta Tsai1,2, Hui-Ling Hsieh2, Shih-Kai Hung2,3, Chi-Fu Zeng2, Ming-Fen Lee2, Po-Hao Lin2, Chia-Yi Lin2, Wei-Chih Li4, Wen-Yen Chiou2,3, Tung-Hsin Wu1 1 Department of Biomedical Imaging and Radiological Sciences, National Yang Ming Chiao Tung University, Taipei, Taiwan 2 Department of Radiation Oncology, Dalin Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation, Chiayi, Taiwan 3 School of Medicine, Tzu Chi University, Hualien, Taiwan 4 Departments of Radiation Oncology, Taichung Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation, Taichung, Taiwan Radiol Oncol 2024; 58(2): 289-299. Received 4 November 2023 Accepted 3 January 2024 Correspondence to: Tung-Hsin Wu, Ph.D., Department of Biomedical Imaging and Radiological Sciences, National Yang Ming Chiao Tung University, No. 155, Sec. 2, Linong St., Beitou Dist., Taipei City 112304, Taiwan, E-mail: tung@ym.edu.tw; Tel: (886) 02-28201095 and Wen-Yen Chiou, MD, Ph.D., Department of Radiation Oncology, Dalin Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation, No. 2, Ming Sheng Road, Dalin Town, Chiayi, 622401, Taiwan, E-mail: cwyncku@gmail.com; Tel: (886) 05-2648000 extension 5695. Disclosure: No potential conflicts of interest were disclosed. This is an open access article distributed under the terms of the CC-BY license (https://creativecommons.org/licenses/by/4.0/). Background. Craniospinal irradiation (CSI) poses a challenge to treatment planning due to the large target, field junction, and multiple organs at risk (OARs) involved. The aim of this study was to evaluate the performance of knowl- edge-based planning (KBP) in CSI by comparing original manual plans (MP), KBP RapidPlan initial plans (RPI), and KBP RapidPlan final plans (RPF), which received further re-optimization to meet the dose constraints. Patients and methods. Dose distributions in the target were evaluated in terms of coverage, mean dose, conform- ity index (CI), and homogeneity index (HI). The dosimetric results of OARs, planning time, and monitor unit (MU) were evaluated. Results. All MP and RPF plans met the plan goals, and 89.36% of RPI plans met the plan goals. The Wilcoxon tests showed comparable target coverage, CI, and HI for the MP and RPF groups; however, worst plan quality was demon- strated in the RPI plans than in MP and RPF. For the OARs, RPF and RPI groups had better dosimetric results than the MP group (P < 0.05 for optic nerves, eyes, parotid glands, and heart). The planning time was significantly reduced by the KBP from an average of 677.80 min in MP to 227.66 min (P < 0.05) and 307.76 min (P < 0.05) in RPI, and RPF, respectively. MU was not significantly different between these three groups. Conclusions. The KBP can significantly reduce planning time in CSI. Manual re-optimization after the initial KBP is recommended to enhance the plan quality. Key words: knowledge-based planning; RapidPlan; craniospinal irradiation; volumetric modulated arc therapy Introduction Prophylactic or therapeutic craniospinal irradia- tion (CSI) is an option for managing certain pri- mary brain tumors, such as medulloblastoma, or hematologic malignancies.1 Since the maximum field of the linear accelerator is 40 cm by 40 cm, the conventional three-dimensional conformal radia- tion therapy (3D-CRT) techniques for CSI use two opposed lateral craniocervical fields adjoined by two adjacent posterior spinal fields. In convention- al CSI techniques, the fields are matched between Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning290 the lateral and posterior fields, creating over- or underdosage within the spinal cord. To address this issue, 3D-CRT with the moving junction tech- nique2,3, which involves changing different junc- tion locations daily during the treatment course, is an option to blur the dose ununiform effect. The moving junction technique in 3D-CRT re- quires the use of multiple treatment plans, which increases the complexity of treatment planning and daily treatment. Moreover, the CSI moving junction technique can only reduce the dose unu- niform effect but cannot obtain dose homogene- ity as a common treatment. With the development of commercial treatment planning system (TPS), volumetric modulated arc therapy (VMAT) with multi-isocenter optimization4 was introduced. VMAT with 360-degree beams can achieve higher conformity and better dispersion of normal organs compared to conventional 3D-CRT.5,6 The VMAT technique with large field overlaps for low-dose gradient junction could tolerate greater positional shifts while maintaining homogeneous dose.7,8 However, planning CSI using the high-precision VMAT technique is challenging and time-consum- ing for medical physicists due to the long treatment field from the brain to the lumbosacral region, which significantly exceeds the treatment field size of a linear accelerator and involves more than ten organs at risk. Because CSI treatment is relatively rare and only patients with possible malignancy tumor cells seeding in the craniospinal canal re- ceive this treatment, medical physicists in many in- stitutions are unfamiliar with this technique. The rarity of the expertise and complex planning pro- cesses make this process resource-intensive. Knowledge-based planning (KBP) is based on a model of estimating dose-volume histograms (DVHs), which is configured by a library of his- torical treatment plans with the aim of improving planning efficiency.9 In previous studies, KBP has been adopted to treat patients with several cancer types, such as head and neck cancers and pelvic malignancies.10-13 KBP showed improved planning efficiency with well-reserved plan quality in those cancer sites. However, compared to those cancer sites, CSI would require more treatment isocent- ers and patients moving with junction feathering. Moreover, more organs at risk (OARs) needed to be considered in CSI than other treatment sites. Reviewing the literature, previous CSI studies have not compared the plan quality and cost-effec- tiveness of the general manual plan method and the KBP with and without re-optimization. This study aimed to compare the plan qual- ity and efficiency of the original manual plans (MP), KBP initial plans (RPI) (RapidPlanTM, Varian Medical Systems, Palo Alto, USA), and KBP final plans, which received further re-optimization (RPF) for CSI. Patients and methods Ethics statement The Institutional Review Board of the Dalin Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation approved this study (approval num- ber, B10804011-1) and waived the requirement for written informed consent from the patients in- volved because only anonymized images were retrospectively analyzed, and this study did not affect the actual treatments these patients received before. Patients This study retrospectively collected computed tomography (CT) image sets of 38 anonymized adults assessed between 2014 and 2019. All the im- age sets met the requirement of immobilization, supine position, and scan from head to pelvis. The slice thickness and matrix size were 3–5 mm and 512 × 512 voxels, respectively (Figure 1). Target and OAR delineation The clinical target volume (CTV) includes the whole brain and spinal cord, typically extended to the lumbar spine L3 level. Assembled CTV was FIGURE 1. Flowchart of the study design. CSI = craniospinal irradiation; CT = computed tomography; KBP = knowledge-based planning; VMAT = volumetric modulated arc therapy Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning 291 separated into CTV-brain, CTV-spine-superior, and CTV-spine-inferior for the multiple field opti- mization (Figure 2). The PTV-brain was construct- ed by symmetrically extending the CTV-brain by 3 mm and by adding 5 mm margin to the spine area. The maximum and minimum lengths of the CTV were 77.83 cm and 65.40 cm, while those of the PTV were 78.80 cm and 66.38 cm. The mean lengths of the CTV and PTV were 71.15 ± 4.28 cm and 72.23 ± 4.16 cm, respectively. The mean CTV and PTV were 1413.40 ± 162.18 cm3 and 1823.93 ± 192.14 cm3, respectively. For planning evaluation purposes, the PTV-brain, PTV-spine-superior, and PTV-spine-inferior were combined as PTV. Dose prescription The dose prescription was 36 Gy in 18 daily frac- tions. All plans were normalized so that 95% of the PTV received 100% of the prescribed dose. Treatment planning The 38 CT image sets of anonymized adults were imported to Eclipse TPS version 13.6 (Varian Medical Systems, Palo Alto, CA, USA). Overall, six medical physicists were participated in this study. Plans for each patient were reviewed and approved by the same physician. A TrueBeam linear accel- erator (Varian Medical Systems, Palo Alto, CA, USA) equipped with a 120-leaf multileaf collimator was selected. All plans were set as 6 megavoltage for the VMAT technique. Analytical Anisotropic Algorithm dose calculation algorithm, 2.5 mm dose calculation grid, and jaw tracking were used. The mean lateral field size for the brain field is 14.76 ± 0.08 cm, while the average lateral field size for the spine field is 12.42 ± 2.52 cm. These dimen- sions are adjusted to encompass the entire target within a reasonable rotation range. Jaw tracking technique is used to minimize the impact of trans- mission leakage dose to normal organ. The colli- mator rotation angle is set within a range of ± 35 degrees for the head and ± 12 degrees for the spine, according to the physicist’s discretion at the time. The whole target length was more than 100 cm, whereas the maximum single-field size of a linear accelerator at the isocenter is 40 × 40 cm. Therefore, multiple fields and three isocenters were required. The PTV-brain used two full arcs, with the iso- center positioned at the center of the brain. For the PTV-spine, two or four partial arcs were used on the PTV-spine-superior, and PTV-spine-inferior to avoid the 60–120-degree and 240–300-degree direction for arm sparing. For the sake of clini- cal convenience, the three isocenters were aligned along the same X-axis (left-right). The spine iso- center shared the same X and Y coordinates, differ- ing only along the Z-axis (craniocaudal) (Figure 2). A total of 38 MPs were generated for the 38 patients, with 23 MPs used to train the RapidPlan (RP) model, and 15 MPs used for validation and comparison (Figure 1). Using RP, 15 RP initial plans (RPI) were generated without manual modi- fication, on which we performed further manual re-optimization to generate 15 RP final plans (RPF). Finally, we compared the following three plan groups: MP, RPI, and RPF. Knowledge-based planning The RapidPlan is a commercial KBP program inte- grated within the Eclipse TPS. The KBP program references a library of previously clinically accept- ed treatment plans. It analyzes the geometric and dosimetric features, such as structure sets, field geometry, dose matrices and plan prescriptions of those plans to train a statistical model. This mod- A B C FIGURE 2. Example of the target and field setup. (A) The arrangement of the brain field (dotted lines), spine-superior field (solid lines), spine-inferior field (dashed lines), and their isocenters. Each field overlaps at least 5 cm for the low-dose gradient junction. (B) Full arc was used on the brain field. (C) Partial arc was used in the spinal fields for arm sparing. Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning292 el is then used to predict an achievable range of DVHs and generate dose-volume objectives for a new plan. RapidPlan algorithm The RapidPlan algorithm comprises two main components: model configuration and DVH es- timation. The model configuration component is responsible for setting up new DVH estimation models, which are subsequently utilized in the DVH estimation component to generate estimates for an individual plan. The model configuration component encompasses two distinct phases: data extraction and model training. On the other hand, the DVH estimation component encompasses the phases of estimation generation and objective gen- eration. The minimum requirement of data extraction and model training was 20 plans with their targets and OARs. Among the 20 randomly selected plans for model training, the right lens of three plans were too small to evaluate. Therefore, we added three more plans to meet the training requirement. The model training phase within the DVH es- timation algorithm is dedicated to the creation of DVH estimation models. The estimation genera- tion phase calculates for each supported structure the same metrics that were calculated during the data extraction of the DVH estimation model, ex- cept for the DVH. Once the estimation generation phase has derived the upper and lower bound DVHs, the optimization objectives placement phase translates them into optimization objectives. Plan quality, planning time, and monitor unit comparison There were 27 dosimetric goals of irradiated fields and OARs were evaluated for the three groups among 15 patients. One patient had previously undergone thyroidectomy, and his thyroid dose could not be evaluated. This resulted in a total of 404 items being calculated for model evaluation. Dosimetric characteristics, such as V95, V100, V107, Dmean, Dmax, and D2 of CTV, and PTV, were evalu- ated. In addition, conformity index (CI) and homo- geneity index (HI) of the targets and dose gradi- ent (Rx%) were compared.14 The Radiation Therapy Oncology Group (RTOG) criteria define CI values to be between 1.0 and 2.0 in accordance with the protocol, 2.0 to 2.5 and 0.9 to 1.0 as a minor devia- tion, and > 2.5 and < 0.9 as a major deviation from the protocol. The CI was defined as a ratio between the volume covered by the reference isodose (36 Gy) and the target volume, as in Equation [1]. [1] where VRI = Reference isodose volume and TV = target volume. The HI is the ratio between maximum isodo- se and reference isodose. The formula of HI was shown as Equation [2]. The ideal value is 1, which increases as the plan becomes less homogeneous. [2] Where Imax = maximum isodose in the target and RI = reference isodose. The dose gradient (Rx%) formula is given below: [3] where Vx% = percentage of isodose volume, and TV = target volume. The pre-optimization, optimization, and re- optimization planning times were compared. The pre-optimization time included OARs contouring and field setup, and the re-optimization time was the time of further optimization and calculation until the plan was satisfied. Average monitor units (MUs) were also evaluated. Statistical analysis The Wilcoxon test was used to compare the differ- ences between the three groups. The differences in the dose coverage, mean dose of the targets, and OARs were compared with a 95% confidence in- terval. All tests were two-sided. A p value of < 0.05 was considered statistically significant. SPSS sta- tistical package (version 17; SPSS Inc., Chicago, IL) was used for all statistical analysis. Results Target coverage and OAR sparing Table 1 shows the dosimetric results of targets. For the V100, V107, Dmax, and D2 of the CTV, both MP and RPF groups were significantly better than RPI (P < 0.01). MP and RPF in most subjects were not significantly different, except for V95. For PTV, the V100 was normalized to 95% prescribed dose for all three groups, MP, RPI, and RPF. MP and RPF groups had significantly better V107, Dmax, D2, and HI than did the RPI group (P < 0.01). The MP group had a worse CI than the other groups. In addition, among 13 compared parameters (Table 1), the RPI had worse results in 84.62% (11/13) parameters Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning 293 compared to the MP and RPF groups, which had the best results in 30.77% (4/13) and 61.53% (8/13) parameters, respectively. The value of HI was the same in the MP and RPF groups. Furthermore, there were 14 OARs and 20 evalu- ation parameters for these OARs (Table 2). RPF and RPI had better dosimetric results than MP for the Dmean of optic nerves, parotid glands, heart, and esophagus, and Dmax of eyes (all P < 0.05). The RPF group was significantly better than the RPI group in 11 parameters (P ≤ 0.01); no parameter in the RPF group was worse than any parameter in the RPI group. RPF had comparable results to the MP group in the other OARs including, brain, brain stem, chiasma, lens, thyroid, lungs, liver, and kidneys. In conclusion, when comparing the three groups, except the heart V40, which was 0% in all these three groups, the MP and RPI groups obtained the worst results in 63.16% (12/19) and 36.84% (7/19) OAR parameters, respectively. On the contrary, the RPF group had 73.68% (14/19) OAR parameters that were superior or equal to the other two groups. Overall, the RPF group achieved superior or equal best results in 71.88% (23/32) of the 32 evalu- ation parameters of the targets (13) and OARs (19), which excluding the PTV V100% and heart V40Gy, because the volumes were the same in all three groups. In this study, we evaluated the quality of the treatment plans for three groups of 15 patients each. We used 27 parameters to evaluate each plan, for a total of 404 parameters, due to one pa- tient who did not have a thyroid gland. We did not include the parameters CTV V107%, CTV Dmean, CTV Dmax, PTV V107%, PTV Dmean, CI, and HI in the evalu- ation because they did not have specific goal val- ues. The plan quality pass rate of the MP and RPF groups was 100% (404/404) according to the plan goals of targets and OARs. The RPI group pass rate was 89.36% (361/404). When evaluating the failures of the RPI group, although no patient in the RPI group passed the CTV V100 goal of 99%, the minimum and median values of RPI CTV V100 were 97.83% and 98.44%, respectively, and both the CTV V95 and the PTV V95 of RPI group reached the goals. TABLE 1. Dosimetric comparison between manual plans, RapidPlan initial, and RapidPlan final Parameters Goals Results P value MP RPI RPF MP vs. RPI MP vs. RPF RPI vs. RPF CTV V95 [%] > 99 99.99 ± 0.03 99.98 ± 0.03 99.97 ± 0.03 0.36 0.03* 0.09 V100 [%] > 99 99.20 ± 0.17 98.37 ± 0.33 99.37 ± 0.23 < 0.01** 0.07 < 0.01** V107 [%] Minimize 0.62 ± 0.59 2.94 ± 4.33 0.46 ± 0.66 < 0.01** 0.16 < 0.01** Dmean [Gy] 36 37.23 ± 0.18 37.31 ± 0.21 37.22 ± 0.24 0.07 0.87 0.13 Dmax [Gy] Minimize 39.38 ± 0.40 40.38 ± 0.57 39.42 ± 0.41 < 0.01** 0.78 < 0.01** D2 [%] < 107 106.12 ± 0.73 106.95 ± 0.96 105.72 ± 0.84 < 0.01** 0.19 < 0.01** PTV V95 [%] > 98 99.68 ± 0.15 99.55 ± 0.23 99.24 ± 0.32 0.03* < 0.01** < 0.01** V100 [%] = 95 95.00 ± 0.00 95.00 ± 0.00 95.00 ± 0.00 - - - V107 [%] Minimize 0.62 ± 0.57 3.01 ± 4.12 0.44 ± 0.61 < 0.01** 0.17 < 0.01** Dmean [Gy] 36 37.10 ± 0.16 37.22 ± 0.18 37.08 ± 0.20 0.05 0.73 0.01* Dmax [%] < 112 109.99 ± 1.17 112.89 ± 1.78 110.17 ± 1.14 < 0.01** 0.57 < 0.01** D2 [%] < 107 106.09 ± 0.73 107.00 ± 0.93 105.71 ± 0.80 < 0.01** 0.21 < 0.01** CI 1 0.98 ± 0.01 1.01 ± 0.01 1.00 ± 0.01 < 0.01** < 0.01** 0.01* HI 1 1.10 ± 0.01 1.13 ± 0.02 1.10 ± 0.01 < 0.01** 0.57 < 0.01** CI = conformity index; CTV = clinical target volume; Dx = minimum dose received by the hottest x% volume; HI = homogeneity index; MP = manual plan; PTV = planning target volume; RPI = RapidPlan initial; RPF = RapidPlan final; Vx = volume receiving at least x dose; * = P < 0.05; ** = P < 0.01 Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning294 TABLE 2. Dosimetric goals and results for organs at risk OAR parameters Goals Results P value MP RPI RPF MP vs. RPI MP vs. RPF RPI vs. RPF Brain Dmax [Gy] < 60 39.34 ± 0.39 40.24 ± 0.61 39.28 ± 0.37 < 0.01** 0.46 < 0.01** Brain stem Dmax [Gy] < 54 38.48 ± 0.41 39.03 ± 0.46 38.51 ± 0.28 < 0.01** 0.91 < 0.01** Chiasm Dmean [Gy] < 50 37.15 ± 0.35 37.10 ± 0.32 36.95 ± 0.32 0.69 0.07 0.06 Dmax [Gy] < 55 38.13 ± 0.40 38.73 ± 0.55 38.20 ± 0.26 0.01* 0.46 < 0.01** Optic nerves Dmean [Gy] < 50 27.61 ± 3.40 22.90 ± 2.39 22.42 ± 2.29 < 0.01** < 0.01** 0.05 Dmax [Gy] < 55 37.13 ± 0.69 36.36 ± 1.72 36.15 ± 1.39 0.13 0.02* 0.13 Eyes Dmax [Gy] < 50 25.55 ± 3.57 22.52 ± 3.83 21.60 ± 3.86 0.02* 0.01* 0.01* Lens Dmax [Gy] < 10 8.40 ± 0.68 8.87 ± 1.00 8.10 ± 0.55 0.11 0.21 < 0.01** Parotid glands Dmean [Gy] < 25 7.38 ± 2.52 5.16 ± 0.39 4.95 ± 0.39 < 0.01** < 0.01** < 0.01** Spinal cord Dmax [Gy] < 50 38.93 ± 0.51 39.81 ± 0.67 39.04 ± 0.56 < 0.01** < 0.01** < 0.01** Thyroid Dmax [Gy] < 45 17.23 ± 4.04 16.68 ± 2.13 16.38 ± 2.04 0.59 0.36 0.07 Lungs Dmean [Gy] < 13 4.63 ± 0.30 4.95 ± 0.43 4.63 ± 0.24 0.03* 0.73 < 0.01** V20Gy [%] < 22 0.06 ± 0.11 0.04 ± 0.08 0.03 ± 0.04 0.64 0.44 0.33 V5Gy [%] < 42 36.48 ± 2.88 42.77 ± 5.62 37.11 ± 2.87 0.01* 0.69 < 0.01** Heart Dmean [Gy] < 10 6.76 ± 1.47 5.53 ± 0.82 5.70 ± 0.93 0.01* 0.02* 0.33 V40Gy [%] < 3 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 - - - V18Gy [%] < 5 0.04 ± 0.10 0.01 ± 0.03 0.01 ± 0.02 0.31 0.23 0.41 Esophagus Dmean [Gy] < 34 14.34 ± 1.62 13.23 ± 1.63 13.30 ± 1.74 0.01* < 0.01** 0.96 Liver Dmean [Gy] < 30 4.82 ± 0.94 4.57 ± 0.73 4.45 ± 0.70 0.61 0.33 < 0.01** Kidneys Dmean [Gy] < 18 2.81 ± 1.17 2.47 ± 0.46 2.38 ± 0.43 0.96 0.73 < 0.01** OAR = organ at risk; MP = manual plan; RPI = RapidPlan initial; RPF = RapidPlan final; Vx = volume receiving at least x dose; * = P < 0.05; ** = P < 0.01 Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning 295 The pass rates of CTV D2, PTV Dmax, and PTV D2, for the RPI group, were 66.67% (10/15), 33.33% (5/15), and 66.67% (10/15), respectively. In addition, in the OAR, the lens Dmax and lungs V5 of the RPI group did not meet the goals. The pass rate of the lens Dmax was 93.33% (14/15) for the RPI group. In one RPI plan, the lens Dmax was 10.98 Gy > 10 Gy. Lastly, the RPI lungs V5 pass rate was 53.33% (8/15). Table 3 shows the mean dose of the 9 OARs. The highest OARs Dmean of the optic nerves, eyes, pa- rotid glands, thyroid, heart, liver, and kidneys; and lens and lungs in these three groups were obtained in the MP group (78%, 7/9) and RPI group (22%, 2/9), respectively. The lowest OARs Dmean were mostly in the RPF group (89%, 8/9). Comparing RPI and MP, RPF and RPI, and RPF and MP groups, the RPI group significantly reduced the doses of optic nerves, eyes, parotid glands, and heart than the MP group; the RPF group further significantly reduced the doses of eyes, lenses, parotid glands, thyroid, lungs, liver, and kidneys than the RPI group (P ≤ 0.05); and RPF significantly reduced the doses of optic nerves, eyes, parotid glands, thyroid, and heart, respectively than the MP group (P < 0.05). In the low-dose region of normal tissue, we em- ployed R50%, R30%, and R10% as dose gradient indica- tors. The values for MP, RPI, and RPF at R50% were 2.27 ± 0.13, 2.26 ± 0.16, and 2.26 ± 0.14, respectively. For R30%, the values were 3.96 ± 0.31, 3.95 ± 0.32, and 3.94 ± 0.37, respectively. The corresponding values for R10% were 10.15 ± 1.93, 10.08 ± 1.69, and 10.00 ± 1.74. There were no statistically significant differ- ences among the three groups (P > 0.05). TABLE 3. The mean dose of the OARs outside the targets contours Organ Mean dose P value MP RPI RPF MP vs. RPI MP vs. RPF RPI vs. RPF Optic nerves 27.61 ± 3.40 22.90 ± 2.39 22.42 ± 2.29 < 0.01** < 0.01** 0.05 Eyes 12.11 ± 1.99 10.16 ± 0.55 9.83 ± 0.74 0.01* 0.01* 0.02* Lens 7.09 ± 0.67 7.21 ± 0.52 6.78 ± 0.39 0.43 0.16 < 0.01** Parotid glands 7.38 ± 2.52 5.16 ± 0.39 4.95 ± 0.39 < 0.01** < 0.01** < 0.01** Thyroid 10.41 ± 3.37 9.00 ± 1.94 8.51 ± 2.07 0.06 0.04* < 0.01** Lungs 4.63 ± 0.30 4.95 ± 0.43 4.63 ± 0.24 0.03* 0.73 < 0.01** Heart 6.76 ± 1.47 5.53 ± 0.82 5.70 ± 0.93 0.01* 0.02* 0.33 Liver 4.82 ± 0.94 4.57 ± 0.73 4.45 ± 0.70 0.61 0.33 < 0.01** Kidneys 2.81 ± 1.17 2.47 ± 0.46 2.38 ± 0.43 0.96 0.73 < 0.01** Bold type = the highest Dmean in the three groups; MP = manual plan; RPI = RapidPlan initial; RPF = RapidPlan final; Underline mark = the lowest Dmean in the three groups; * = P < 0.05; ** = P < 0.01 FIGURE 3. (A) Population-averaged dose-volume histogram (DVH) for all organs at risk and targets. (B) The population-averaged DVH for targets only. CTV = clinical target volume; MP = manual optimization plan; PTV = planning target volumes; RPI = RapidPlan initial; RPF = final RapidPlan after manual re-optimization A B Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning296 Figure 3A showed the population-averaged DVH of targets and OARs. In the DVH, the doses of optic nerves, eyes, lens, parotid glands, thyroid, liv- er, and kidneys in RPF or RPI were lower than those in MP. Furthermore, the DVH of RPF OARs was bet- ter than those of RPI OARs. Figure 3B shows the tar- gets coverage of CTV and PTV. In the shoulder part of the DVH, with the 95% volume of targets, the MP and RPI groups had the same targets coverage, while the RPF group had a slightly better 95% vol- ume dose coverage than the other two groups. The DVH tail part, the high dose in 5% volume, showed that the RPI had the highest dose in the craniospi- nal area. The population-averaged DVH showed that the RPF group had the best targets coverage, homogenous targets dose distribution, and OAR dose avoidance among these three groups. Treatment planning time The pre-optimization time was the same in all three groups (146 minutes, Figure 4). The optimi- zation process took a significantly longer time in the MP group than in the RPI and RPF groups with 111.45, 81.68, and 81.68 minutes (P < 0.05), respec- tively. The re-optimization time in the MP was significantly longer than in the RPF group (420.36 versus 85.13 minutes, P < 0.05). There was no re- optimization in the RPI group. Overall, the entire planning time was longer in the MP group than in the RPI (677.80 versus 227.66 minutes, P < 0.05) and RPF (677.80 versus 307.76 minutes, P < 0.05) groups. The total planning time-saving rates (saved plan- ning time) of RPI and RPF were 66.41% (450.14 min- utes) and 54.59% (370.04 minutes), respectively, compared to the MP group. MU comparison The average MU values with one standard de- viation of MP, RPI, and RPF groups were 935.24 ± 128.44, 1013.22 ± 114.92, and 1026.46 ± 149.43, re- spectively, with no significant difference between these three groups (all P > 0.05). Discussion Our research discovered that by utilizing 23 plans to develop the KBP model in combination with RP and re-optimization in CSI, we were able to sig- nificantly shorten the planning time by half and enhance plan quality. Incorporating more patients in the model librar- ies for model training have a possibility to lead to fewer outliers and more consistent plan quality.15-17 However, the application of the CSI technique in clinical practice is not common in most hospitals. In this study, because CSI treatment is relatively ra- re, we searched databases covering the previous 6 years and found only 38 CT image sets. The Varian accelerator company recommended a minimum of 20 to 25 treatment plans in training set for a spe- cific target. According to the study by Jim P. Tol et al.18, Increasing the number of plans used in model training was found to produce comparable results. Based on recommendations, previous experience, and the limited availability of clinical CSI cases, we used 23 plans to complete the model training and compared them with 15 manual plans. The traditional CSI used patient prone position to reduce the OARs radiation dose via simple two lateral opposed and posterior-anterior (PA) fields. However, this technique can create dose unu- niform in the field junction area. The commonly encountered pediatric CSI typically requires two fields and one junction to achieve coverage. This study aims to validate whether KBP can perform effectively in more complex scenarios, utilizing adult CSI as a test case. We used the VMAT tech- nique to disperse the radiation dose in OARs and enhance the homogeneity of the targets dose. The VMAT technique delivers radiation from all an- gles, which causes it to be attenuated as it passes FIGURE 4. Comparison of the planning time for MP, RPI, and RPF. The error bar represents one standard deviation. MP = manual optimization plan; RPI = RapidPlan initial; RPF = final RapidPlan after manual re- optimization Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning 297 through the couch. Our medical physicist compen- sated for this effect by calculating the attenuation of the couch.19 Furthermore, cone beam computed tomography ensured an accurate treatment loca- tion. Therefore, in this study, all treatment plans were designed using the supine position, which could make patients more comfortable, relaxed, and stable during treatment.3,20 Although the plan parameter pass rate of RPI was only 89.36%, the RPI target coverage of mini- mum CTV V95 and PTV V95 values were ≥ 99.90% and ≥ 99.00%, respectively, which were both high- er than 95%, the clinical common plan acceptable criteria.21 Compared with the traditional 3D-CRT technique, by which the high dose area might re- ceive approximately twice the prescribed dose at the field overlapping sites, the highest PTV Dmax in RPI was 115.57% which was much lower than the traditional 3D-CRT technique. For OARs, all 14 plans in RPI achieved the goal (< 10 Gy) except for one plan with lens Dmax 10.98 Gy, which did not reach the goal. Table 2 shows that the heart Dmean in RPI was also the lowest of the three groups. Although, Uehara et al. reported that KBP was found clinically unacceptable after a single opti- mization without manual objective constraints in head and neck cancer.22 Most studies in the other body sites, such as gynecological, prostate, and rectal cancers, support that the RP plan would be comparable to the manual plan.23 In our study, the RPI plans were clinically acceptable for CSI and ap- proved by the physician. The DVH distribution is one of the vital plan evaluation tools. The DVH of OARs (Figure 3) showed that most of the OARs in the MP group received higher doses than RPI and RPF, as shown by the Dmean and Dmax in Table 2. In the target DVH (Figure 3B), the RPF group had better 95% volume dose coverage and better performance at reducing high doses than the other two groups. According to our CI results, there was a minor deviation of the target in the MP group; however, RPI or RPF could have achieved the planning goal. Furthermore, HI values in this study show that MP and RPF groups had better homogeneity than did RPI. Previous studies on lung cancer or prostate cancer showed that KBP could reduce the OARs dose23; however, target coverage and dose homogeneity of KBP did not always have better results than the manual plan. Our study on CSI showed that RP improved the plan quality of OARs and that additional re- optimization after initial RP could improve the plan quality, as previous studies showed in other cancer sites.24-27 In terms of cardiac doses, all three plans (MP, RPI, and RPF) exhibited notably low V40Gy and V18Gy values, comfortably below the established cardiac dose constriants. It is pertinent to mention that the mean cardiac dose for RPI was already lower than that for MP. Therefore, the primary focus during the optimization process was not predominantly on further reducing cardiac dose. In the case of RPI, the lungs V5Gy value(42.77 ± 5.62%) surpassed the target threshold of 42%. Subsequently, in the ensuing RPF optimization, concerted efforts were undertaken to amplify the reduction of lungs V5Gy values, resulting in a dose shift towards the heart. Nevertheless, from a statistical perspective, the P-value for the comparison between RPI and RPF exceeded 0.05. In our study, RPI and RPF reduced planning time compared to MP by 66.41% (450.14 minutes) and 54.59% (370.04 minutes), respectively. The result showed that KBP for CSI might save more planning time in complex plans with many OARs than in general cancer sites. Previously, Wells et al.28 reported that KBP could reduce planning time by approximately 30 minutes per breast cancer pa- tient. Visak et al.29 reported that all the RP plans required less than 30 minutes of planning time for lung cancer. Masi et al.30 showed that the time re- quired for the production of the KBP plan was 6–15 minutes, compared to manual planning requiring 30–150 minutes for a commercial TPS and 15–60 minutes after 8 months of commercial TPS usage in prostate cancer. Furthermore, Chatterjee et al.31 showed that the KBP planning time for the multi- form brain glioblastoma was typically 13 minutes for VMAT, compared to the typical 4 hours for the manual planning method. Amaloo et al.32 showed that the total planning time was reduced from 120 minutes to 20 minutes in prostate cancer patients. In a study of nasopharyngeal cancer, Chang et al.33 concluded that the total RP planning time is on- ly about one-fifth that of MP. Similarly, our KBP study for CSI, a very long treatment size from the brain to the lumbosacral area, could effectively re- duce the planning time while improving the plan quality, as shown in previous KBP studies for oth- er cancer sites. Conclusions This study used 23 plans to train the KBP CSI mod- el and investigated the difference between MP and RP for the same patients and found that RP plans after re-optimization could halve the planning Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning298 time and improve plan quality. According to our study result, medical physicists at low CSI patient volume hospitals could efficiently produce CSI plans by the KBP method. Acknowledgements We would like to thank Ms. Feng-Chun Hsu for statistical support, Dr. Liang-Cheng Chen for his advice for clinical aspects of this project. This study was supported by research grants from the Dalin Tzu Chi Hospital (grant number: DTCRD109-I-18). The funders had no role in the study design, data collection and analysis, decision to publish, or the preparation of the manuscript. References 1. Seidel C, Heider S, Hau P, Glasow A, Dietzsch S, Kortmann RD. Radiotherapy in medulloblastoma-evolution of treatment, current concepts and future perspectives. Cancers 2021; 13: 5945. doi: 10.3390/cancers13235945 2. Kiltie AE, Povall JM, Taylor RE. The need for the moving junction in craniospinal irradiation. Br J Radiol 2000; 73: 650-4. doi: 10.1259/ bjr.73.870.10911789 3. Mani KR, Sapru S, Maria Das KJ, Basu A. A supine cranio-spinal irradiation technique using moving field junctions. Pol J Med Phys Eng 2016; 22: 79-83. doi: 10.1515/pjmpe-2016-0014 4. Mancosu P, Cozzi L, Muren LP. Total marrow irradiation for hematopoietic malignancies using volumetric modulated arc therapy: a review of treat- ment planning studies. Phys Imaging Radiat Oncol 2019; 11: 47-53. doi: 10.1016/j.phro.2019.08.001 5. Seravalli E, Bosman M, Lassen-Ramshad Y, Vestergaard A, Oldenburger F, Visser J, et al. Dosimetric comparison of five different techniques for craniospinal irradiation across 15 European centers: analysis on behalf of the SIOP-E-BTG (radiotherapy working group). Acta Oncol 2018; 57: 1240-9. doi: 10.1080/0284186X.2018.1465588 6. Prabhu RS, Dhakal R, Piantino M, Bahar N, Meaders KS, Fasola CE, et al. Volumetric modulated arc therapy (VMAT) craniospinal irradiation (CSI) for children and adults: a practical guide for implementation. Pract Radiat Oncol 2022; 12: e101-e9. doi: 10.1016/j.prro.2021.11.005 7. Sarkar B, Pradhan A. Choice of appropriate beam model and gantry ro- tational angle for low-dose gradient-based craniospinal irradiation using volumetric-modulated arc therapy. J Radiother Pract 2016; 16: 53-64. doi: 10.1017/s146039691600042x 8. Sarkar B, Munshi A, Manikandan A, Roy S, Ganesh T, Mohanti BK, et al. A low gradient junction technique of craniospinal irradiation using volumetric- modulated arc therapy and its advantages over the conventional therapy. Cancer Radiother 2018; 22: 62-72. doi: 10.1016/j.canrad.2017.07.047 9. Hussein M, Heijmen BJM, Verellen D, Nisbet A. Automation in intensity modulated radiotherapy treatment planning-a review of recent innova- tions. Br J Radiol 2018; 91: 20180270. doi: 10.1259/bjr.20180270 10. Ma C, Huang F. Assessment of a knowledge-based RapidPlan model for patients with postoperative cervical cancer. Prec Radiat Oncol 2017; 1: 102- 7. doi: 10.1002/pro6.23 11. Fogliata A, Reggiori G, Stravato A, Lobefalo F, Franzese C, Franceschini D, et al. RapidPlan head and neck model: the objectives and possible clinical benefit. Radiat Oncol 2017; 12: 73. doi: 10.1186/s13014-017-0808-x 12. Hu J, Liu B, Xie W, Zhu J, Yu X, Gu H, et al. Quantitative comparison of knowledge-based and manual intensity modulated radiation therapy plan- ning for nasopharyngeal carcinoma. Front Oncol 2020; 10: 551763. doi: 10.3389/fonc.2020.551763 13. Castriconi R, Fiorino C, Passoni P, Broggi S, Di Muzio NG, Cattaneo GM, et al. Knowledge-based automatic optimization of adaptive early-regression- guided VMAT for rectal cancer. Phys Med 2020; 70: 58-64. doi: 10.1016/j. ejmp.2020.01.016 14. Shaw E, Kline R, Gillin M, Souhami L, Hirschfeld A, Dinapoli R, et al. Radiation Therapy Oncology Group: radiosurgery quality assurance guide- lines. Int J Radiat Oncol Biol Phys 1993; 27: 1231-9. doi: 10.1016/0360- 3016(93)90548-a 15. Boutilier JJ, Craig T, Sharpe MB, Chan TC. Sample size requirements for knowledge-based treatment planning. Med Phys 2016; 43: 1212-21. doi: 10.1118/1.4941363 16. Cagni E, Botti A, Wang Y, Iori M, Petit SF, Heijmen BJM. Pareto-optimal plans as ground truth for validation of a commercial system for knowl- edge-based DVH-prediction. Phys Med 2018; 55: 98-106. doi: 10.1016/j. ejmp.2018.11.002 17. Wang M, Gu H, Hu J, Liang J, Xu S, Qi Z. Evaluation of a highly refined prediction model in knowledge-based volumetric modulated arc therapy planning for cervical cancer. Radiat Oncol 2021; 16: 58. doi: 10.1186/ s13014-021-01783-9 18. Tol JP, Delaney AR, Dahele M, Slotman BJ, Verbakel WF. Evaluation of a knowledge-based planning solution for head and neck cancer. Int J Radiat Oncol Biol Phys 2015; 91: 612-20. doi: 10.1016/j.ijrobp.2014.11.014 19. Yu CY, Chou WT, Liao YJ, Lee JH, Liang JA, Hsu SM. Impact of radiation attenu- ation by a carbon fiber couch on patient dose verification. Sci Rep 2017; 7: 43336. doi: 10.1038/srep43336 20. Sarkar B, Munshi A, Ganesh T, Manikandan A, Mohanti BK. Dosimetric comparison of short and full arc in spinal PTV in volumetric-modulated arc therapy-based craniospinal irradiation. Med Dosim 2020; 45: 1-6. doi: 10.1016/j.meddos.2019.03.003 21. Dietzsch S, Braesigk A, Seidel C, Remmele J, Kitzing R, Schlender T, et al. Pretreatment central quality control for craniospinal irradiation in non- metastatic medulloblastoma: first experiences of the German radiotherapy quality control panel in the SIOP PNET5 MB trial. Strahlenther Onkol 2021; 197: 674-82. doi: 10.1007/s00066-020-01707-8 22. Uehara T, Monzen H, Tamura M, Ishikawa K, Doi H, Nishimura Y. Dose- volume histogram analysis and clinical evaluation of knowledge-based plans with manual objective constraints for pharyngeal cancer. J Radiat Res 2020; 61: 499-505. doi: 10.1093/jrr/rraa021 23. Ge Y, Wu QJ. Knowledge-based planning for intensity-modulated radiation therapy: a review of data-driven approaches. Med Phys 2019; 46: 2760-75. doi: 10.1002/mp.13526 24. Wu H, Jiang F, Yue H, Zhang H, Wang K, Zhang Y. Applying a RapidPlan model trained on a technique and orientation to another: a feasibility and dosimet- ric evaluation. Radiat Oncol 2016; 11: 108. doi: 10.1186/s13014-016-0684-9 25. Castriconi R, Fiorino C, Broggi S, Cozzarini C, Di Muzio N, Calandrino R, et al. Comprehensive Intra-Institution stepping validation of knowledge-based models for automatic plan optimization. Phys Med 2019; 57: 231-7. doi: 10.1016/j.ejmp.2018.12.002 26. Kamima T, Ueda Y, Fukunaga JI, Shimizu Y, Tamura M, Ishikawa K, et al. Multi-institutional evaluation of knowledge-based planning performance of volumetric modulated arc therapy (VMAT) for head and neck cancer. Phys Med 2019; 64: 174-81. doi: 10.1016/j.ejmp.2019.07.004 27. Fogliata A, Cozzi L, Reggiori G, Stravato A, Lobefalo F, Franzese C, et al. RapidPlan knowledge based planning: iterative learning process and model ability to steer planning strategies. Radiat Oncol 2019; 14: 187. doi: 10.1186/s13014-019-1403-0 28. Wells DM, Walrath D, Craighead PS. Improvement in tangential breast plan- ning efficiency using a knowledge-based expert system. Med Dosim 2000; 25: 133-8. doi: 10.1016/s0958-3947(00)00039-x 29. Visak J, McGarry RC, Randall ME, Pokhrel D. Development and clinical vali- dation of a robust knowledge-based planning model for stereotactic body radiotherapy treatment of centrally located lung tumors. J Appl Clin Med Phys 2021; 22: 146-55. doi: 10.1002/acm2.13120 30. Masi K, Archer P, Jackson W, Sun Y, Schipper M, Hamstra D, et al. Knowledge- based treatment planning and its potential role in the transition between treatment planning systems. Med Dosim 2018; 43: 251-7. doi: 10.1016/j. meddos.2017.10.001 Radiol Oncol 2024; 58(2): 289-299. Tsai WT et al. / Performance of knowledge-based treatment planning 299 31. Chatterjee A, Serban M, Abdulkarim B, Panet-Raymond V, Souhami L, Shenouda G, et al. Performance of knowledge-based radiation therapy planning for the glioblastoma disease site. Int J Radiat Oncol Biol Phys 2017; 99: 1021-8. doi: 10.1016/j.ijrobp.2017.07.012 32. Amaloo C, Hayes L, Manning M, Liu H, Wiant D. Can automated treatment plans gain traction in the clinic? J Appl Clin Med Phys 2019; 20: 29-35. doi: 10.1002/acm2.12674 33. Chang ATY, Hung AWM, Cheung FWK, Lee MCH, Chan OSH, Philips H, et al. Comparison of planning quality and efficiency between conventional and knowledge-based algorithms in nasopharyngeal cancer patients using intensity modulated radiation therapy. Int J Radiat Oncol Biol Phys 2016; 95: 981-90. doi: 10.1016/j.ijrobp.2016.02.017