Implementation of a DVH Registry to provide constraints and continuous quality monitoring for pediatric CSI treatment planning

Craniospinal irradiation (CSI) is a complex radiation therapy technique that is used for patients, often children and teenagers / young adults, with tumors that have a propensity to spread throughout the central nervous system such as medulloblastoma. CSI is associated with important long ‐ term side effects, the risk of which may be affected by numerous factors including radiation modality and technique. Lack of standardization for a technique that is used even in larger radiation oncology departments only a few times each year may be one such factor and the current ad hoc manner of planning new CSI patients may be greatly improved by implementing a dose – volume histogram registry (DVHR) to use previous patient data to facilitate prospective constraint guidance for organs at risk. In this work, we implemented a DVHR and used it to provide standardized constraints for CSI planning. Mann – Whitney U tests and mean differences at 95% con ﬁ dence intervals were used to compare two cohorts (pre ‐ and post ‐ DVHR intervention) at speci ﬁ c dosimetric points to determine if observed improvements in standardization were statistically signi ﬁ cant. Through this approach, we have shown that the implementation of dosimetric constraints based on DVHR ‐ derived data helped improve the standardization of pediatric CSI planning at our center. The DVHR also provided guidance for a change in CSI technique, helping to achieve practice standardization across TomoTherapy and IMRT. ﬂ all treatment ﬁ DVHR. treatment as index, plan treatment planning Python DVH cohort MySQL Statistical cohorts performed a interface built JavaScript allowing interactive and dynamic visualization. our case, Cohort dataset, constraints, cohort postintervention dataset.

depend on the age of the patient 11 but overall 60% to 90% of patients develop chronic side effects. [12][13][14][15] Due to the lack of published organ tolerances for CSI planning, physicians and dosimetrists regularly refer to plans of previously treated patients for guidance. Although the Pediatric Normal Tissue Effects in the Clinic (PENTEC 11,16 ) consortium has recently begun a concerted effort to determine organ dose tolerances for pediatric patients, recommendations similar to those provided for adults by the Quantitative Analysis of Normal Tissue Effects in the Clinic (QUANTEC 17,18 ) effort may still be several years away. In the quality improvement project we describe here, we developed a dose-volume histogram registry (DVHR) to review historical dosimetric data from CSI treatment plans and used it to derive institutional dosimetric constraints for OARs. We hypothesized that the implementation and use of our DVHR would provide us with planning constraints and enable continuous quality monitoring of pediatric CSI treatment planning at our institution.

| MATERIALS AND METHODS
We created our DVHR in 2014 to enable the visualization and comparison of DVHs from multiple patients simultaneously. A web interface allows for interactive visualization of DVHs using a dynamic JavaScript charting library (Highcharts, Vik i Sogn, Norway). The DVHR was developed using Django, an open-source web framework based on Python, allowing the use of libraries for data analysis such as pandas, lifelines, and scipy. 19,20 The DVHR front-end uses HTML, CSS, and JavaScript in order to provide an interactive web-based user interface. The Python backend interfaces with a MySQL database.
The DVHR consists of three key elements: a MySQL database to store the DVH data, a web-based user interface, and a series of Python scripts to import, load, and analyze the data. Figure 1 depicts the data flow through the DVHR as used in the presently described project. Patient information is filtered, anonymized, and imported into the database using a custom Python script. Data incorporated from each plan into the DVHR include the prescription dose and fractionation scheme, as well as organ volume (in cm 3 ), absolute and relative DVH points, and the mean, median, maximum, and minimum doses for all targets and OAR structures.
Data within the DVHR may be grouped into "cohorts" and summary statistics may be used to compare and contrast these cohorts.

2.A | Intervention
Beginning in 2014, our institution used the DVHR to establish CSI planning constraints based on our previous planning experience.
Craniospinal irradiation plans were imported from the treatment The plans of nine patients treated for CSI between 2009 and 2014 met these criteria and were accessible for the analysis. Using the dosimetric information of these patients, the population mean and median dose (along with the standard deviation) were calculated for 10 OARs: bilateral lungs, both kidneys, liver, heart, stomach, esophagus, trachea, and thyroid gland. New planning constraints for plan evaluation were established based on the median values and were used by the planners as the initial optimization criteria in plan optimization.
From 2014 onwards, physicians and dosimetrists planned new CSI patients prescribed 36 Gy in 20 fractions using the DVHRderived constraints. By early 2020, nine new patients had been treated and their data incorporated into the DVHR, at which point we evaluated the impact of the DVHR-derived constraints on clinical practice.

2.B | Evaluation
We undertook a pre-post analysis to determine the usefulness of the DVHR as a tool to monitor CSI treatment planning at our center.
Two cohorts of CSI plans (using the standard prescription of 36 Gy in 20 fractions) were examined: ten plans from before the DVHRderived constraints were implemented in the clinic (preconstraints, incorporating the nine plans used to derive the constraints and one additional plan) and nine plans from after the implementation (postconstraints).
Plans included in this study were treated with two different radi-

2.C | Dose reduction examination
With the aim of visually observing pre-post changes for specific dosimetric points, the mean DVH and its standard uncertainty was plotted for each OAR-cohort combination. This allowed us to quickly and qualitatively determine across the board if there was any reduction in the dose delivered to the OARs due to the intervention of the DVHR-derived constraints.
In order to determine if use of the DVHR-derived constraints led to a change in practice, the central tendencies (mean and median) and spread (standard deviation and interquartile range) of the D mean values of each OAR were compared before and after the intervention, as any reductions postintervention could indicate an improvement in practice. D mean was selected for this exercise as it provides a good summary of the dose to the whole structure and is widely applicable to OARs. Violin plots were used to evaluate changes in the median and interquartile ranges of the D mean value of the population for each OAR. This method of plotting numeric data is similar to box plots with the advantage of also showing the probability density of the data at each dose value. 23,24 Due to the change in treatment modality, it was important to investigate if any changes in OAR sparing and treatment planning standardization after 2014 were due to the modality change or due to the use of the DVHR-derived constraints. For this reason, we F I G . 1. Schematic showing the data flow of the DVH registry. Using a Python Script, DVH data from all patients meeting the selection criteria in the Eclipse treatment planning system were accessed via the Eclipse API, anonymized, filtered, renamed, and inserted into the DVHR. The internal serial number of the patient within the treatment planning system was used as an index, providing external anonymization while preserving a link to the original plan in the treatment planning system if needed subsequently. A series of Python scripts were used to extract DVH data for display or cohort median values from the MySQL database. Statistical comparisons between patient cohorts were performed using a web interface built with JavaScript allowing interactive and dynamic visualization. In our case, Cohort 1 corresponded to the preintervention dataset, which provided the constraints, and cohort 2 corresponded to the postintervention dataset.
used timeline plots to examine the change in OAR dose over time in order to determine if any observed improvements began before the switch to IMRT.

2.D | Mann-Whitney U tests
Mann-Whitney U tests were used to assess the pre-and post-DVHR cohorts at each of the specific dosimetric points to determine if the observed changes in summary statistics were statistically significant. 25,26 We selected this nonparametric test of the null hypothesis (i.e., that there is no difference between cohorts) over others as our data are not normally distributed and our sample size is small. 27 Eight tests, corresponding to V 5Gy , V 10Gy , V 15Gy , V 20Gy , D mean , D median , D min , and D max , comparing the pre-and postconstraints cohorts, were performed.

2.E | Means and confidence intervals
In general, sample size plays an impactful role in the results obtained from hypothesis testing evaluations. Therefore, Mann-Whitney U tests, although insightful, may not be sufficient to adequately demonstrate the significance of changes in practice. [28][29][30] Because of this, we also compared the pre-and post-DVHR cohorts with respect to their population mean scores for all dosimetric points (V 5Gy , V 10Gy , V 15Gy , V 20Gy , D mean , D median , D min , and D max ) and calculated the confidence intervals of the differences between them. The evaluation of the difference in means between cohorts at the 95% confidence interval was computed using the appropriate t distribution for the selected confidence level and the standard uncertainty of the point estimate. 31 The execution of all statistical analyses was conducted using a custom-written Python script incorporating a two-tailed (P < 0.05) significance level.

3.A | Intervention
The DVH constraints derived in 2014 are shown in Table 1 for all ten OARs (the lungs, kidneys, liver, heart, stomach, esophagus, trachea, and thyroid gland) examined in this study.

3.B | Evaluation
All individual DVHs for each patient OAR, before and after implementation of the planning constraints, are shown in Fig. 3. Large structures such as the heart, lungs, kidneys, liver, and stomach displayed a clear reduction in dose in the postconstraints period compared to the preconstraints period. Other structures, particularly smaller structures close to the target volume, had less or no reductions in dose after implementation of the DVHR.

3.C | Dose reduction examination
The population-mean DVHs of the pre-and postconstraints cohorts are shown in Fig. 4. The heart, right lung, liver, and kidneys had the most dramatic decrease in their mean DVHs postconstraints, whereas the left lung, esophagus, stomach, and trachea demonstrated less obvious or minimal reductions. The thyroid gland was the only OAR that displayed an increased dose after the intervention, unlike all other organs which displayed either a lower or similar population-mean DVH in the postconstraints cohort.
Following the implementation of the constraints in 2014, the interquartile ranges of seven of ten OARs decreased for the D mean parameter: the right lung, kidneys, liver, stomach, thyroid, and trachea (Fig. 5). The right kidney and trachea had the most pronounced changes in their interquartile ranges, decreasing from 2.7 to 0.61 Gy and from 5.2 to 4.2 Gy, respectively. In contrast, the esophagus, heart, and left lung increased their interquartile ranges postconstraints by 2.2, 0.7, and 0.4 Gy, respectively. This was also consistent with the spreads seen visually in the individual DVHs (Fig. 3).
Except for the trachea and the thyroid gland, all OARs had a reduced population mean and median of the D mean value postintervention (Fig. 5). The most dramatic decrease in population median D mean was for the liver, left and right kidneys, which reduced by 29%, 43%, and 47%, respectively.
Supplementary material presented at the end of this manuscript contains detailed information regarding the population median of D mean and the interquartile ranges for all OARs.
The mean values of D mean for each year of the study are depicted in Fig. 6 as temporal trends for all OARs. The mean value of D mean reduced after the constraints were introduced but before the change in technique for all structures except for the thyroid gland.

3.D | Mann-Whitney U tests
According to the Mann-Whitney U tests (results shown in Fig. 7), the introduction of the new dosimetric constraints led to statistically significant reductions in dose to the heart, right lung, both kidneys, and liver. This can be observed from the significant decreases in V 5Gy , V 10Gy , D mean , D median , and D min for these structures. Although there were some reductions in V 5Gy , V 10Gy , V 15Gy , V 20Gy for the left lung, esophagus, and trachea postconstraints, these reductions were not statistically significant. Overall, dose to OARs remained similar or was reduced nonsignificantly for most structures. However, V 20Gy to the liver, D max to the esophagus and trachea, and V 10Gy , V 15Gy, V 20Gy, D mean , D median , and D max to the thyroid gland, all increased after constraints were implemented, although those increases were not statistically significant. However, some improvements in practice standardization were observable prior to the change for all OARs, except the thyroid gland, using timeline trends that show the trend of mean dose over the years. Also, the reduction in dose variability is more likely attributable to consistent planning than to a change in technique.

3.E | Means and confidence intervals
We can also attest that the use of consistent DVHR-derived con- As future work, quantitative treatment outcomes will be extracted from the medical records and used to assess the clinical impact of the reduced dose to the OARs. Use of the DVHR will also be applied to other radiotherapy techniques and treatment sites that lack monitoring of practice.

CONF LICT OF I NTEREST
There is no conflict of interest to disclose.