Dose Prediction Model for Duodenum Sparing With a Biodegradable Hydrogel Spacer for Pancreatic Cancer Radiation Therapy.

PURPOSE
We previously have shown the feasibility of duodenum sparing using a biodegradable hydrogel spacer in pancreatic cancer radiation therapy. In this study, we propose an overlap volume histogram (OVH) prediction model to select patients who might benefit from hydrogel placement and to predict the hydrogel spacing required to achieve clinical constraints.


METHODS AND MATERIALS
OVH metrics for the duodenum were collected from the stereotactic body radiation therapy plans of 232 patients with unresectable pancreatic cancer (33 Gy in 5 fractions). OVH metrics L9cc and L3cc were defined as the tumor volume expansion distance at which 9 cm3 and 3 cm3 volumes of the duodenum overlap with tumor. D9cc and D3cc of the duodenum were defined as the dose-volume histogram dose to 9 cm3 and 3 cm3, respectively, of the duodenum. Prediction models were established by linear regression between Lx and Dx, where x = 3 cm3 and 9 cm3. OVH thresholds were obtained for predicting the target spacer thickness. The accuracy of the prediction model was then evaluated using treatment plans on pre-and post-hydrogel injection computed tomography scans from 2 cadaver specimens and 6 patients with previously treated locally advanced pancreatic cancer with simulated spacer.


RESULTS
Linear regression analysis showed a significant correlation between Lx and Dx (r2 = 0.51 and 0.51 for L3cc-D3cc and L9cc-D9cc, respectively; both P < .01). The OVH thresholds were Lˆ3cc = 7 mm and Lˆ9cc = 13 mm. The observed planning doses D3cc and D9cc of duodenum from pre-and post-hydrogel injection computed tomography scans of cadaver specimens and clinical patients with simulated spacer using predicted target spacer thickness were within the OVH model prediction range.


CONCLUSION
Our model may predict which patients require placement of a hydrogel spacer before stereotactic body radiation therapy to meet predefined dose constraints. Furthermore, by predicting the required target hydrogel thickness, the spacer injection can be better guided to improve efficacy.


INTRODUCTION
Pancreatic cancer is the third most common cause of cancer-related death and remains the most devastating cancer with a 5-year relative survival rate of merely 8% in the United States [1]. With current therapeutic approaches, only a minority of patients presenting with resectable disease have a chance for long-term survival. One-third of patients will present with borderline resectable or locally advanced pancreatic cancer (BR/LAPC). Although these patients still present with localized disease, median overall survival (OS) is a poor 9 to 15 months [2][3][4][5]. The mainstay treatment for BR/LAPC patients is single-or multi-agent chemotherapy or chemoradiation (CRT) in sequence with chemotherapy.
Until recently, the proximity and inherent radiosensitivity of the surrounding gastrointestinal tracts, particularly the stomach and duodenum, were considered an impasse in the feasibility of further dose-intensification using stereotactic body radiation therapy (SBRT) [6][7][8][9][10] and intensity modulated radiation therapy (IMRT) [11]. In other sites in body, however, physical separation between the target and adjacent organ-at-risk (OAR), has evolved into an effective method of reducing the dose to the OAR and the toxicity of dose-escalated RT. This has been most successfully demonstrated in the use of a hydrogel for spacing the rectum from the prostate in the treatment of prostate cancer [12][13][14][15]. In a phase-three study of dose-escalated IMRT for prostate cancer, physical separation of the rectum from the prostate using a hydrogel spacer significantly decreased the dose delivered to the rectum and resulted in reduced toxicity and improved quality of life for patients undergoing spacer placement [14]. MA) to separate the head of the pancreas (HOP) from the duodenum via endoscopic ultrasound (EUS) guidance in a cadaveric model [16]. TraceIT consists of a hydrogel paste that creates a bleb of particles at the needle tip upon injection. The bleb maintains its threedimensional structure for 3 months and is fully absorbed after approximately 7 months based on manufacturer research and development investigations. We demonstrated successful separation of the HOP and duodenum using the hydrogel as a spacer, establishing the promise of the technique to potentially overcome this physical barrier to dose-escalation. We also reported a minimum separation distance of 8 mm would achieve significant dose reduction to the duodenum across all modeled cases in a simulation study-a distance of separation that was achievable with the minimally invasive EUS-guided spacer injection technique [16]. Our group is now proceeding with a clinical trial to attempt the EUS-guided hydrogel spacer placement between the HOP and duodenum in patients with BR/LAPC pancreatic cancer without duodenal invasion. Prior to embarking on this trial, however, we aim to establish a prediction model for determining a patient-specific target separation distance to better guide the hydrogel injection process for each patient's individual anatomy, accounting for the position of the duodenum, pancreatic tumor, target prescription dose, and clinical dose constraints.
Overlap volume histogram (OVH) has been previously proposed to quantify the geometrical relationship between a target and OAR. By comparing the OVH of a new patient with prior treated patients, the previously used dose volume histogram (DVH) constraints can be retrieved from patient library to determine if the DVH constraints in the new patient are achievable [17]. This approach can also be used to correlate OVH and DVH data to establish OVH prediction models for semi-automated [18][19][20] and automated treatment planning [21].
Based on these experiences, OVH prediction modeling has the potential to accurately predict the doses administered to different structures and to estimate target separation distances to achieve particular planning objectives and constraints. Here, we aim to propose and evaluate an ideal workflow that uses a novel OVH prediction model to guide the selection of patients with BR/LAPC without duodenal invasion who may benefit from placement of a hydrogel spacer and to predict a minimum target separation distance using a hydrogel spacer between the HOP and duodenum prior to radiation therapy to achieve clinical dose constraints.

Patient inclusion criteria for the OVH prediction model
Following an Internal Review Board (IRB) approval, 232 patients with BR/LAPC who were treated with SBRT at our institution from 2011 to 2016, were included in a database to generate an OVH prediction model.

Planning methods
Patient simulation, contouring techniques, and prescription dose were consistent across all patients included in the OVH prediction model. Patients were positioned supine with an alpha cradle immobilization and underwent a simulation computed tomography (CT) scan (Philips Brilliance Big Bore CT, 2 mm slice thickness, 120 kVp, 200 mA, 60 cm field of view) as previously described [7]. The gross tumor volume (GTV) included the primary tumor, delineated using the simulation CT, fusing the diagnostic CT and positron-emission tomography/CT (PET/CT) scan when available. The tumor motion as a result of the breathing cycle was assessed on four-dimensional (4D) CT at time of simulation. The GTV was manually contoured and was expanded to account for motion using an ITV (internal target volume) or patients were treated using Active Breathing Coordinator (ABC) (Elekta, Stockholm, Sweden) if feasible for the patient. A 2-3 mm planning target volume (PTV) expansion was applied. OARs including the duodenum, stomach, liver, kidneys, and spinal cord were manually contoured (Pinnacle, Philips Radiation Oncology Systems, Milpitas, CA) and dedicated contours of the regions of the duodenum, stomach, and small bowel within 1 cm superior and inferior extent of the PTV (prox-duod, prox-stomach, prox-small bowel) were contoured [7].

Developing an OVH prediction model
Based on the SBRT protocol passing criteria prox-duod V15Gy<9cc and V20Gy<3cc, we selected the specific points used in the OVH prediction model as follows: L 3cc and L 9cc = Radial expansion of the PTV in millimeters to encompass 3cc and 9cc, respectively, of the prox-duod volume within the expanded PTV. The resolution of this measurement is 1.5 mm, and it is chosen to be slightly higher than our minimal CT slice thickness which is at 1 mm. D 3cc and D 9cc = DVH dose (Gy) at 3cc and 9cc volume of the prox-duod. We established a model using L 3cc and L 9cc from the OVH to predict D 3cc and D 9cc ,, respectively using the same hypothesis proposed in previous studies [23]. It assumes a direct relationship between the L x point in OVH and the D x point in the DVH. This prediction was analyzed with respect to the clinical criteria of D3cc<20Gy and D 9cc <15Gy.
The L 3cc , L 9cc , D 3cc , and D 9cc were calculated for 232 patients in the Pinnacle treatment planning system using an in-house developed software program with Pinnacle scripts and structured query language (SQL). The OVH prediction model was then created by applying linear regression and Pearson correlation tests to examine the relationship between L3cc and D3cc, and L9cc and D9cc.
The values of the upper bound of the 95% confidence interval of the intersection of D 3cc =20 Gy and L 3cc and the intersection of D 9cc =15 Gy and L 9cc were defined as the OVH thresholds L 3cc and L 9cc , respectively. The OVH thresholds represent the L3cc and L9cc for which the predicted prox-duod dose met the planning criteria D 3cc <20Gy and D 9cc <15Gy in 95% of patients.
OVH thresholds L 3cc and L 9cc were used to calculate the target spacer thickness, ΔL, between the prox-duod and the PTV necessary to meet the planning criteria. Using an OVH prediction model and OVH thresholds L 3cc and L 9cc , we proposed a patient selection workflow that could be used to evaluate if a new patient who will require radiation therapy would benefit from spacer placement as shown in Figure 1. The workflow starts following simulation CT and contouring of the target and OARs. OVH parameters L 3cc and L 9cc are extracted as the input to the OVH prediction model to calculate the predicted D 3cc and D 9cc values. If the predicted values meet the clinical constraint, the patient can go directly to SBRT planning and treatment. Otherwise, the patient would be recommended to undergo hydrogel spacer placement with a target spacer thickness ΔL, followed by repeat simulation, and subsequent SBRT planning and treatment.

Evaluating the accuracy of the OVH prediction model and proposed workflow
To evaluate the accuracy of the OVH prediction model and proposed workflow, we used two types of CT datasets for SBRT treatment planning. The first test set evaluated only the prediction accuracy of the OVH prediction model and the second test set evaluated the accuracy of the OVH model incorporated into the proposed workflow.
The first test dataset included the CT scans of two human cadaveric specimens (Specimen 1 and Specimen 2) pre-and post-hydrogel spacer placement on an IRB approved study. The hydrogel spacer, synthesized as iodinated polyethylene glycol microparticles, was injected under EUS-guidance by an experienced gastroenterologist (XX) to separate the HOP and the 2 nd and 3 rd portion of the duodenum as previously described [16]. A mock 2 cm spherical tumor in the HOP was contoured as the GTV with an additional 2mm expansion to the PTV. OARs in the cadaver study were contoured with identical criteria as in clinical practice and as described in above [16,24]. The OVH data points (L 3cc , L 9cc ) were calculated using our in-house software for both the pre-and post-spacer CTs of Specimen 1 and Specimen 2. The predicted D 3cc and D 9cc were calculated using the OVH prediction model for each CT based on its values of L 3cc and Lg cc . The accuracy of the OVH prediction model was assessed by comparing the predicted D 3cc and D gcc to the corresponding actual SBRT planning dose D 3cc and D 9cc obtained on pre-and post-spacer CT scans for Specimen 1 and Specimen 2.
The second test dataset included 6 randomly selected patients with BR/LAPC who underwent SBRT at our institution and were excluded from the cohort used to build the OVH prediction model. We followed the workflow as in Figure 1 to demonstrate the patient selection process and used the simulated spacer to validate our proposed OVH prediction model and OVH thresholds (L 3cc and L 9cc ).
For each of the 6 patients, the OVH parameters L 3cc and L 9cc were extracted from the CT scan and the predicted dose to the prox-duod D 3cc and D 9cc were calculated based on the proposed OVH prediction model. When constraints for D 3cc and D 9cc were not met, then the

Author Manuscript
Author Manuscript Author Manuscript

Author Manuscript
OVH thresholds (L 3cc and L 9cc ) were used to calculate the target spacer thickness ΔL as follows: Based on the ΔL value, the HOP-duodenum displacement was simulated in these patients by shifting the prox-duod contour away from the PTV manually by a magnitude of ΔL. SBRT plans were re-optimized after introducing the simulated space between the prox-duod and the PTV, with the same planning objectives and constraints in 100 iterations. By comparing the predicted ranges of D 3cc and D 9cc to the corresponding resultant D 3cc and D 9cc on the post-simulated spacer plans, the accuracy of the OVH prediction model, the proposed workflow (Figure 1), and ΔL was evaluated.

Statistical analysis
The relationships between L 3cc and D 3cc and L 9cc and D 9cc were examined by the Pearson correlation test and linear regression. Validation of the OVH model was considered successful if the resultant D 3cc and D 9cc of the simulation spaced prox-duod fell within the model predicted ranges of both metrics.

OVH prediction model
The correlations between the OVH variables (L 3cc , L 9cc ) and dose to the prox-duod (D 3cc , D 9cc ), as shown in Figure 2, were determined as follows: OVH thresholds L 3cc = 7mm and L 9cc = 13mm, as defined above, were obtained. The OVH prediction model was composed of these two functions and their corresponding thresholds.

Evaluating the accuracy of the OVH prediction model for radiation treatment planning
A. Test dataset #1: Cadaveric specimen CTs with injected spacer-OVH metrics (L 3cc , L 9cc ) from SBRT plans pre-and post-spacer placement for Specimen 1 and Specimen 2 were used to predict the dose to the prox-duod (D 3cc , D 9cc ) in a total of 4 plans. Predicted and observed dosimetric results are shown in Table 1. The mean thickness of the spacer was 1.1 cm (range 0.9-1.2 cm) and 0.9 cm (range 0.8-1.1 cm) for Specimen 1 and 2 on post-hydrogel injection CT, respectively. Optimized pre-spacer SBRT plans failed to achieve clinical constraints for D3cc in Specimen 1 and 2. With increased separation due to the injected spacer, D 3cc decreased from 20.2Gy and 20.7Gy to 16.9Gy and 17.5Gy for Specimen 1 and 2, respectively. Figure 3 depicts the SBRT plan, prox-duod OVH, and PTV and prox-duod DVH pre-and post-spacer placement for Specimen 1. The overlap between the PTV volume and the prox-duod decreased from 1.12cc to 0.01cc ( Figure 3C) and PTV coverage by the prescription dose (33Gy) improved from 93% to 96% ( Figure 3D) with spacer placement.

B. Test dataset #2: Clinical patients CTs with simulated spacer-OVH
variables (L 3cc , L 9cc ) from each of the 6 sample clinical patients (A-F) were used to predict the dose to the prox-duod (D 3cc , D 9cc ). Predicted and observed dosimetric results are shown in Table 2. As all 6 sample patients had undergone SBRT at our institution, the observed D 3cc and D gcc achieved or closely approached the pre-specified constraints. In addition, the ΔL gcc values were all greater than Δ L 3cc . Consequently, the predicted target hydrogel separation distances ΔL were set as ΔL gcc . On the baseline sample clinical plans, D 3cc values of Patient A (20.1Gy) and Patient C (20.6Gy) and D gcc value of Patient F (15.5Gy) slightly exceeded the clinical protocol constraints (D 3cc <20Gy, D gcc <15Gy). Following simulation of the spacer placement, the radiation doses received by the prox-duod were significantly reduced in all cases. On average, D 3cc and D gcc of prox-duod decreased by 15% and 5%, respectively.
The baseline and simulated post-spacer SBRT plan of sample Patient E is shown in Figure  4A and B. Following the predicted separation distance (ΔL of Patient E = 10 mm), the overlap between the PTV volume and the prox-duod decreased from 2.4cc to <0.1cc ( Figure  4C). Simulated spacer placement improved PTV coverage by the prescription dose (33Gy) from 73% to 93%, and the V15Gy and V20Gy of the prox-duod decreased by 48% and 84%, respectively ( Figure 4D).

DISCUSSION
We propose an OVH prediction model and workflow to guide the selection of patients with BR/LAPC without duodenal invasion who may benefit from placement of a hydrogel spacer to achieve predetermined clinical dose constraints. For patients flagged by the model where spacer placement would be recommended to achieve a clinically acceptable plan, the prediction model furthermore can be used to predict a minimum target spacer thickness needed to separate the duodenum from the PTV in the HOP. This work provides valuable insights into the dosimetric advantages of hydrogel spacer injection to achieve a patientspecific target separation distance that accounts for individual anatomy of the duodenum and pancreatic target and desired final radiotherapy plan characteristics, including the potential for dose-escalated radiation therapy.
By correlating the changes of DVH and OVH from spacer injection, our results suggest that OVH has promising potential for guiding both patient section and the desired spacer thickness for the procedure of using hydrogel injection to space the HOP and duodenum. Because an injected hydrogel spacer between duodenum and HOP results in a change in the relative geometrical relationship of these two structures, we demonstrate that OVH as a metric could be used to establish the link between radiation dose and required separation between duodenum and HOP.
OVH data has previously been applied to predict objectives and constraints of OARs to set the foundation for semi-automatic and automatic treatment planning. The majority of experience in applying OVH data to radiotherapy planning is with head and neck [18,20] and prostate cancer planning [19,21]. XXXX and colleagues, introduced OVH as an efficient method to generate achievable DVH objectives and offered a tool to predict clinically achievable doses prior to planning [20]. Their study not only demonstrates the effectiveness of this method in head and neck radiotherapy planning, but also illustrates that IMRT planning no longer needs to be a time-consuming process. By combining OVH and a database of previous treatment plans, the investigators herald the possibility of automated treatment planning. Until now, however, "automation" still required certain skilled manual inputs to modify the planning parameters in training cycles [25].
OVH modeling has also previously been applied in dose prediction in external beam prostate radiotherapy with hydrogel spacer injection [21]. XXXX and colleagues used OVH metrics predict dose to the rectum to serve as the dose objective for automated treatment planning [21]. The authors predicted rectal sparing with spacer placement using OVH metrics L 20 % (distance of expansion of a PTV corresponding to 20% of rectal volume overlap with the expanded PTV) and D 20 % (dose at 20% volume on rectal DVH).
We expanded the application of these principles to the radiotherapeutic management of BR/ LAPC patients. In our study, we proposed using absolute OVH metrics (L 3cc , L 9cc , D 3cc , and D 9cc ) rather than relative OVH metrics given the absolute volume dose constraints used at our institution for hypofractionated SBRT. Unlike lung, parotid and other parallel OARs, ablation of a small volume in a serial OAR such as the duodenum has significant clinical consequence. Thus, absolute OVH metrics are more appropriate to use in this clinical scenario.
As a result of the work developing and testing the OVH prediction model described here, we are able to propose a workflow ( Figure 1) to guide prudent selection of patients and, when relevant, a patient-specific target spacer thickness to be achieved with hydrogel injection. This is one more tool that can be applied in radiation oncology to practice precision medicine, customizing the use of the hydrogel spacer device appropriately and optimally based on the context of each patient's tumor, normal tissue anatomy, and prescription dose. Our proposed workflow also facilitates the urgent need for dose-escalated radiation therapy using the advantages of spacer injection, as dose-escalation to the PTV has previously been limited by the adjacent radiosensitive duodenum. The OVH model generated in this study will be prospectively applied to patients undergoing hydrogel spacer placement in a clinical trial to spare the dose-limiting duodenum.
Several limitations of this study should also be considered. As the 232 patients involved in the model database were participants in previous clinical trials, their treatment plans were created with a large variation of PTV coverage, caused by compromises for protecting the duodenum and stomach. Thus, the predicted range was relatively broad (D 3cc : ±2.82Gy, D 9cc : ±2.55Gy). Additionally, all patients were treated at our institution, which may lead to bias in radiation treatment methods. The thresholds determined from this model were conservative. L 3cc and L 9cc lower than these thresholds might still meet the clinical constraints for the duodenum. Nevertheless, as future clinical SBRT patients will be treated with this hydrogel injection technique in different institutions, more OVH and DVH data will become available for further optimizing our OVH prediction model. This will ultimately result in the determination of more accurate thresholds and better prediction ranges.

CONCLUSION
In this study, we proposed an OVH prediction model to guide the selection of patients with pancreatic cancer who may benefit from placement of a hydrogel spacer with a target required thickness to achieve predetermined clinical dose constraints for duodenum sparing in BR/LAPC patients. Furthermore, this work provides valuable insights into the dosimetric advantages of hydrogel spacing, which may be exploited for future investigations of doseescalated radiation therapy for patients with pancreatic cancer.  OVH data points L 3cc and L 2cc (A) and L 9cc and L 3cc (B) were plotted for all 232 patients (star). Linear regression was used to analyze the correlation between the metrics with 95% prediction range (dash line) and 95% confidence interval (grey shading). The L 3cc value at the intersection of the D 3cc =20Gy with the upper bound of the 95% confidence interval of the linear regression line of L3cc and D 3cc was calculated and defined as L 3cc . Similarly, the L 9cc value at the intersection of the D 9cc =15Gy with the upper bound of the 95% confidence interval of the linear regression line of L 9cc and D 9cc was calculated and defined as L 9cc . The vertical lines represent the OVH thresholds L 3cc and L 9cc .  The PTV (orange), duodenum (green), 33Gy prescription isodose line (red), 20Gy isodose line (yellow), and 15Gy isodose line (blue) are shown on pre-spacer (A) and post-spacer (B) Specimen 1 SBRT plans. The resultant changes in the prox-duod OVH metric (C) and PTV coverage (D) pre-(solid) and post-spacer placement (dashed) are shown. The overlap between the PTV volume and the prox-duod decreased from 1.12cc to 0.01 cc (arrow, Figure  3C).  The orange contours in color wash represent the PTV, and the green contours are the proxduod. The red, yellow and blue isodose lines represent 33Gy (prescription dose), 20Gy and 15Gy, respectively. The arrows in A and B represent the decrease in the low dose region.
The arrow in C represents the decreased overlap between prox-duod and PTV. The resultant changes in the prox-duod OVH metric (C) and prox-duod and PTV DVH metric (D) pre-(solid) and post-spacer placement (dashed) are shown.