Associations between voxel-level accumulated dose and rectal toxicity in prostate radiotherapy.

BACKGROUND AND
PURPOSE: Associations between dose and rectal toxicity in prostate radiotherapy are generally poorly understood. Evaluating spatial dose distributions to the rectal wall (RW) may lead to improvements in dose-toxicity modelling by incorporating geometric information, masked by dose-volume histograms. Furthermore, predictive power may be strengthened by incorporating the effects of interfraction motion into delivered dose calculations.Here we interrogate 3D dose distributions for patients with and without toxicity to identify rectal subregions at risk (SRR), and compare the discriminatory ability of planned and delivered dose.
MATERIAL AND METHODS: Daily delivered dose to the rectum was calculated using image guidance scans, and accumulated at the voxel level using biomechanical finite element modelling. SRRs were statistically determined for rectal bleeding, proctitis, faecal incontinence and stool frequency from a training set (n = 139), and tested on a validation set (n = 47).
RESULTS: SRR patterns differed per endpoint. Analysing dose to SRRs improved discriminative ability with respect to the full RW for three of four endpoints. Training set AUC and OR analysis produced stronger toxicity associations from accumulated dose than planned dose. For rectal bleeding in particular, accumulated dose to the SRR (AUC 0.76) improved upon dose-toxicity associations derived from planned dose to the RW (AUC 0.63). However, validation results could not be considered significant.
CONCLUSIONS: Voxel-level analysis of dose to the RW revealed SRRs associated with rectal toxicity, suggesting non-homogeneous intra-organ radiosensitivity. Incorporating spatial features of accumulated delivered dose improved dose-toxicity associations. This may be an important tool for adaptive radiotherapy in the future.

Introduction

Rectal toxicity remains a clinical issue in prostate radiotherapy. Existing toxicity prediction models may no longer be appropriate for the modern high-dose gradient complex treatments currently being delivered, as the historic data used to derive constraints for dose-limiting organs were based on 3D conformal techniques [1], [2]. The underlying mechanisms and pathophysiology of rectal toxicity remain relatively poorly understood [3], [4].

Dose-volume histogram (DVH) constraints are commonly applied in radiotherapy treatment planning to achieve an optimal compromise between tumour coverage and healthy organ sparing. However, it is widely acknowledged that DVH parameters are limited by their lack of spatial dose information [5], [6], [7]. Studies into more appropriate dose descriptors for the rectal wall (RW) have generally been focused on parameterisation of dose surface maps (DSMs) where the RW is virtually cut and unfolded to a normalised 2D map so that geometric dose information can be preserved [5], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16].

An advancing area of research, which complements interrogation of localised dose features, is voxelwise dose-toxicity analysis [17], [18]. Investigating associations between voxel-level dose and toxicity can improve the accuracy in identifying heterogeneous areas of heightened dose sensitivity, or rectal subregions at risk (SRRs) [18]. One solution offering voxel-resolution anatomical modelling and 3D tracking of organ motion, is biomechanical finite element (FE) analysis [19], [20], [21], [22], [23], [24]. FE modelling applies constitutive equations to anatomical simulations, and unlike the DSM method, is not restricted to normalised in-plane expansion. Additionally, by comparison with image intensity-based deformable registration tools which can struggle at large deformations and are generally also limited to planar expansions [25], FE offers superior spatial functionality.

An overarching limitation and recommendation of studies examining dose-toxicity associations in prostate radiotherapy, has been that motion-inclusive dosimetric data are required for a more complete understanding of the underlying mechanisms [4]. Radiation treatments to the pelvic region are particularly susceptible to interfraction motion which can lead to large deviations between planned dose and that actually delivered. This introduces an inherent uncertainty in dose-toxicity modelling based on planned dose.

Previous work [16] presented the calculation and accumulation of delivered dose to DSMs of the RW. The 2D DSMs were resampled such that the number of pixels spanning the width was equal to the rectal length from the planning scan [7], [8]. However, when applied to dose accumulation, the DSM approach was limited in that modelling of rectal expansion and deformation was in-plane only and uniformly normalised about the circumference of the rectum. The motivation for implementing biomechanical FE modelling into the VoxTox workflow was to provide a more anatomically representative basis for voxel-level accumulation of delivered dose [26]. FE models have previously been developed for the rectum in prostate radiotherapy [22], [23], but to date, dose accumulation studies have focused on the liver [21].

The aim of this study was to determine rectal SRRs from planned and accumulated FE-DSMs, and investigate associations with toxicity.

Material and methods

VoxTox study design & patient information

The VoxTox research programme is an observational study linking radiation dose to toxicity outcomes [27]. It received approval from the National Research Ethics Service (NRES) Committee East of England (13/EE/0008) in February 2013 and is part of the UK Clinical Research Network Study Portfolio (UK CRN ID 13716). Appropriate consent was obtained from all participants.

Data were analysed for a set of 186 prostate cancer patients recruited to the VoxTox study. This ‘consolidation’ cohort consisted of patients prescribed 74 Gy in 37 fractions (n = 110, 59%) and 60 Gy in 20 fractions (n = 76, 41%), treated with TomoTherapy® (Accuray, Sunnyvale, CA), and whose baseline characteristics and follow-up data were formally and prospectively recorded. Planning dose constraints were applied following the CHHiP trial protocol [28], [29]. Doses for patients prescribed 60 Gy in 20 fractions were radiobiologically converted to the equivalent dose in 37 fractions using an alpha/beta ratio of 2.1 Gy based on an analysis of rectal toxicity in the CHHiP trial [28], [29]. Equivalent levels of cumulative toxicity incidence were assumed between the two prescription arms.

All patients were treated with image-guided intensity modulated radiotherapy (IG-IMRT), with a daily megavoltage computed tomography (MVCT) scan acquired immediately prior to treatment delivery for online target localisation [30]. According to local imaging protocols, if rectal dilation was deemed excessive, remedial action was taken prior to delivering the radiation treatment. Kilovoltage (kV) planning scans were manually contoured, and the rectum was defined from the rectosigmoid junction to the most inferior slice containing both ischial tuberosities [31]. For daily MVCT scans, the rectum was identified using a locally developed autosegmentation system based on the Chan-Vese algorithm [32]. Manual contouring of the 4174 MVCT scans (over 62,000 slices) in this study would not have been practicable, and the autosegmentation system has previously been shown to fall within an acceptable range for intra- and inter-observer variability [33]. Dose was calculated directly from the MVCT images (and recalculated for kVCT images) using CheckTomo, an independent ray-tracing dose calculation algorithm [34], [35]. Where the MVCT field of view was shorter than the kVCT, missing data were substituted using planned dose [16].

Toxicity data

Toxicity data were prospectively collected using study-specific electronic clinical reporting forms. These were designed for robust collection of toxicity data, and the raw data were used to populate recognised systems. Mapping rules were externally validated. Here, we focused on toxicity endpoints affecting the patient’s quality of life, commonly reported in the literature [6], [7], [10], [11], [13], [15], [18], [36], and present results for rectal bleeding G2, proctitis G2, faecal incontinence G1, all Common Terminology Criteria for Adverse Events (CTCAE) v4.03 [37] and stool frequency G1, Late Effects on Normal Tissue: Subjective, Objective, Management (LENT/SOM) scale [38]. Patients were included where a minimum of two years’ follow-up data were available, with complete late toxicity follow-up history recorded at 6, 12, and 24 months. Cumulative incidence at 2 years was investigated, and event rates are shown in Table 1.

Table 1

Two-year toxicity incidence rates for full cohort, training (75%) and validation (25%) sets, stratified by prescription regime. RB = rectal bleeding, proc = proctitis, incont = faecal incontinence, freq = stool frequency.

Toxicity endpoint	Total cohort n (%)	74 Gy n (%)	60 Gy n (%)	Training n (%)	74 Gy n (%)	60 Gy n (%)	Validation n (%)	74 Gy n (%)	60 Gy n (%)
	186 (100)	110 (59)	76 (41)	139 (75)	82 (59)	57 (41)	47 (25)	28 (60)	19 (40)
RB	21 (11)	10 (5)	11 (6)	16 (12)	8 (6)	8 (6)	5 (11)	2 (4)	3 (6)
Proc.	27 (15)	13 (7)	14 (8)	20 (14)	9 (6)	11 (8)	7 (15)	4 (9)	3 (6)
Incont.	32 (17)	21 (11)	11 (6)	24 (17)	16 (12)	8 (6)	8 (17)	5 (11)	3 (6)
Freq.	45 (24)	28 (15)	17 (9)	34 (24)	21 (15)	13 (9)	11 (23)	7 (15)	4 (9)

Training and validation sets

Patients were split into training (n = 139, 75%), and validation (n = 47, 25%) sets. In order to retain the same ratios of toxicity incidence as the whole cohort, the composition of patients within training and validation sets was uniquely generated for each individual toxicity endpoint. The data were non-randomly split by stratifying first by fractionation, then by toxicity incidence. Patients were then randomised into either training or validation set. This maintained toxicity rates between each group as much as possible, as shown in Table 1.

Finite element modelling

A physical-based rectal simulation was created in the Abaqus Computer Aided Engineering (CAE) FE Analysis environment (Dassault Systèmes) using 30 elements around the rectal circumference and 80 along the rectal length. This was approximately equivalent to the width dimensions used in the previous DSM analysis [16], but with increased resolution along the length of the rectum in order to improve the accuracy of tracking extra-planar motion. The model was able to deform and expand in 3D according to material properties, loading and boundary conditions. The boundary conditions were defined by the rectal contours, manually defined on kVCT or autosegmented on MVCT images. Displacement loading was applied such that the model expanded based on the shape and position of the contours, rather than pressure loading where defining parameters can be difficult due to challenges in experimental verification [26]. Further details are provided in Supplementary Material 1. When each simulation was complete, the dose was determined in MATLAB (MathWorks®, Natick, MA) at each voxel by interpolation of the 3D dose matrix calculated in CheckTomo. The use of a common starting point for each simulation allowed voxels to be tracked, and dose histories accumulated. This facilitated like-for-like comparisons between patients, as well as between planned and delivered dose. The resulting dose surface maps generated using FE modelling will be termed FE-DSMs.

Identifying rectal subregions at risk

Rectal SRRs were determined using the training set for each toxicity endpoint following a similar approach to that of Dréan et al. [18]. For each voxel of the FE-DSM, a two-sample t-test for equality of means was performed in MATLAB. The two groups were defined by patients with and without toxicity and the null hypothesis was that there was no difference in mean voxel dose between groups. Voxels were identified in areas where the dose was significantly higher for patients with toxicity than those without. Thresholds were determined based on an unadjusted p-value <0.05, as well as p-values adjusted for false discovery rates using the Benjamini-Hochberg procedure [39]. Where the p-value at each voxel was less than the threshold, the dose difference between patients with and without toxicity was considered significant, and the null hypothesis was rejected. From the resulting voxelwise p-value map, a binary mask was generated whereby a value of 1 was assigned when the null hypothesis was rejected, and a value of 0 was assigned otherwise. Post-processing was performed on the binary mask to remove clusters of fewer than 10 pixels (unless clusters would have been connected had it not been for the posterior splitting of the FE-DSM for visualisation in 2D), to smooth sharp edges and discontinuities using a least-squares smoothing filter (Savitzky-Golay filter in the MATLAB image processing toolbox), and to fill in any remaining holes. This affected fewer than 5% of pixels for all endpoints other than rectal bleeding which was up to 10% due to the proportionately larger SRR and detected holes. The final SRRs for both unadjusted and adjusted thresholds are overlaid on the p-value maps for each endpoint shown in Fig. 1 (for binary masks, see Supplementary Material 2).

Fig. 1

Map of p-values for rectal bleeding (RB), proctitis (proc), stool frequency (freq), and faecal incontinence (incont), indicating regions of significant dose difference between patients with and without toxicity for planned and accumulated dose to the rectal wall. White dashed line represents SRRs where p < 0.05, red dashed line indicates adjusted p-value SRR, where applicable. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

The SRR binary masks were applied to each patient’s planned and accumulated FE-DSM such that only the dose to voxels from the non-zero regions were considered. The equivalent uniform dose (EUD) [40] was then calculated (using a = 11.1 [2]) based on the dose to voxels within these subregions. EUD of the full RW was also calculated, as this was previously found to be associated with rectal bleeding grade 1 and proctitis using 2D DSMs[16]. Receiver operator characteristic (ROC) analysis was performed in SPSS (IBM, 23.0.0.2) and the area under the curve (AUC) was reported with 95% confidence intervals (CIs). Odds ratios (ORs) were also calculated using logistic regression. The analysis was repeated by applying the SRR binary masks to the validation set (which had been kept blind to the process of generating SRRs). Validation AUC and ORs were calculated in order to compare with training set results.

Results

Localised dose response in subregions at risk

The p-value maps shown in Fig. 1 indicate the SRRs of the RW where significant dose difference were observed between patients with and without toxicity. Spatial patterns of localised dose response differed per endpoint. The relative proportions of SRR area (from unadjusted p-value maps) with respect to the full RW were: rectal bleeding 47/56%, proctitis 22/22%, faecal incontinence 3/6%, stool frequency 2/11%, from planned/accumulated dose. For rectal bleeding, the SRRs spanned the circumference of the inferior RW and most of the superior RW for both planned and accumulated FE-DSMs, with additional expansion into the mid-right-posterior RW on the accumulated FE-DSM. The SRRs associated with rectal bleeding were much larger in area than those found for other toxicities. For proctitis, SRRs were observed from the mid-to-inferior posterior RW and connected by a band spanning the circumference inferiorly, common to both planned and accumulated FE-DSMs. For faecal incontinence, a cluster at the inferior right-lateral RW was common to both planned and accumulated FE-DSMs, with an additional mid-posterior to mid-right cluster on the accumulated FE-DSM. For stool frequency, a single cluster was observed in the right-posterior mid-RW for planned dose, and for accumulated dose, a larger cluster spanning the full circumference was located in the mid-upper RW. By comparison, adjusted p-value maps resulted in reduced, more focussed SRR regions for rectal bleeding and proctitis, and did not produce a result for faecal incontinence or stool frequency. Results discussed below refer to unadjusted p-value SRRs unless otherwise specified.

Dose-toxicity associations

When comparing EUD of the full RW, accumulated dose was lower for all four endpoints by an average of 1 Gy (range −0.9, −1.1; p < 0.001) with respect to planned dose (Table 2). The difference in EUD between patients with and without toxicity was greater for accumulated dose for three of four endpoints, but differences could only be considered significant for rectal bleeding. In contrast, when focusing on SRR regions, EUD was greater for accumulated dose than planned dose by an average of 10.6 Gy (range 2.6, 19.5; p < 0.001). The difference in EUD between patients with and without toxicity was greater for SRR than RW. When splitting SRR EUD by toxicity group, the difference was greater from planned dose for three of four endpoints, but the standard deviation was reduced for accumulated dose compared with planned dose by an average of 4.6 Gy (range −2.0, −11.4). However, the only EUD differences considered significant between toxicity groups were for accumulated dose to SRRs for rectal bleeding and faecal incontinence.

Table 2

Equivalent uniform dose (EUD) comparison between full rectal wall (RW) and subregions at risk (SRR) for planned and accumulated dose, split by toxicity groups for the training set. Standard deviations (SD) and EUD differences () are presented between patients with and without toxicity alongside corresponding p-values. Differences between planned and accumulated dose for RW and SRR had p0.001 for all endpoints. RB = rectal bleeding, proc = proctitis, incont = faecal incontinence, freq = stool frequency.

Endpoint	Dose	EUDRW	SD(EUDRW)	Δ_tox(EUDRW)	p-value	EUDSRR	SD(EUDSRR)	Δ_tox(EUDSRR)	p-value
RB	Planned	60.9	2.5	1.2	0.006	51.4	7.2	2.7	0.2
	Accumulated	59.8	3.0	1.7	0.003	54.0	4.7	3.8	0.002
Proc	Planned	61.0	2.3	0.8	0.2	39.5	9.2	3.8	0.1
	Accumulated	60.0	2.7	0.9	0.2	43.1	7.2	3.2	0.06
Incont	Planned	61.0	2.4	0.2	0.8	31.9	17.4	3.0	0.4
	Accumulated	60.0	2.9	0.2	0.8	51.4	6.0	2.7	0.05
Freq	Planned	60.9	2.4	0.4	0.4	42.0	8.1	2.3	0.2
	Accumulated	60.0	2.9	0.7	0.2	58.7	5.7	1.7	0.1

AUCs and ORs for SRR dose-toxicity analysis (with corresponding 95% CIs) are shown in Fig. 2. Results for full RW and adjusted p-values are included alongside a summary of these results in Table 3 (95% CIs included in Supplementary material 3). When comparing AUCs for the full RW, only rectal bleeding was found to have AUC > 0.6, and accumulated dose was more strongly associated than planned dose. For planned dose, there was no advantage to using EUD to SRRs rather than the full RW in terms of discriminative ability. For rectal bleeding, proctitis, and faecal incontinence, AUCs from accumulated SRRs were greater than those from the RW by 7%, 4%, and 12%, respectively. Results for stool frequency could not be considered significant. For SRRs, AUCs were greater when investigating associations from accumulated dose with respect to planned dose (rectal bleeding +15%, proctitis +5%, faecal incontinence +10%). For these endpoints, accumulated dose results had AUC > 0.6 with min 95% CI > 0.5, indicating stronger associations than corresponding planned dose results (where min 95% CI < 0.5). OR results indicated that accumulated dose was more strongly associated with rectal bleeding and incontinence than planned dose, by +20% and +8%, respectively. However, the lower bounds of the 95% CIs were <1 for proctitis, stool frequency, and all planned dose results. For the two endpoints yielding SRRs from p-value maps after adjustment for false discovery rates, rectal bleeding AUC for accumulated dose improved by 4% and for planned dose was unchanged, whereas for proctitis the AUC reduced by 2–3% for both planned and accumulated dose. Differences were not significant compared to unadjusted results.

Fig. 2

(a) Area under the receiver operator characteristic curve (AUC), and (b) odds ratios (OR), with corresponding 95% confidence intervals (CI), for the training set (n = 139). RB = rectal bleeding, proc = proctitis, incont = faecal incontinence, freq = stool frequency.

Table 3

Dose-toxicity analysis for planned and accumulated equivalent uniform dose (EUD) to the full rectal wall (RW) and rectal subregion at risk (SRR). Area under the ROC curve (AUC) and odds ratios (OR) were calculated to compare training (t) and validation (v) sets. Results for SRRs based on p-values adjusted for false discovery rates were also included (AUC). Confidence intervals (CI) for SRR results are shown in Fig. 2, Fig. 3, and are included for all results in Supplementary material 3. RB = rectal bleeding, proc = proctitis, incont = faecal incontinence, freq = stool frequency.

Endpoint	FE-DSM	AUCRW(t)	AUCSRR(t)	AUCFDR(t)	AUCSRR(v)	ORSRR(t)	ORFDR(t)	ORSRR(v)
RB	Planned	0.63	0.61	0.61	0.58	1.06	1.05	1.04
	Accumulated	0.69	0.76	0.80	0.63	1.26	1.28	1.12
Proc	Planned	0.61	0.60	0.58	0.45	1.04	1.14	0.95
	Accumulated	0.61	0.65	0.63	0.39	1.07	1.06	0.93
Incont	Planned	0.54	0.53	–	0.45	1.01	–	0.99
	Accumulated	0.51	0.63	–	0.49	1.09	–	1.02
Freq	Planned	0.58	0.58	–	0.49	1.04	–	0.97
	Accumulated	0.60	0.60	–	0.58	1.06	–	0.98

Fig. 3

(a) Area under the receiver operator characteristic curve (AUC). and (b) odds ratios (OR), with corresponding 95% confidence intervals (CI), for the validation set (n = 47). RB = rectal bleeding, proc = proctitis, incont = faecal incontinence, freq = stool frequency.

Validation results, as shown in Fig. 3, were inherently lower with larger CIs than corresponding training results as expected due to the smaller sample size. Due to the widths of the CIs, the validation results could not be considered sufficiently significant to verify the findings from the training set.

Discussion

The investigation of associations between localised dose to rectal SRRs and toxicity is an area of increasing interest. In the current study, subregions of the RW were identified by statistical evaluation of voxel-level dose differences between FE-DSMs of patients with and without toxicity. The addition of biomechanical FE modelling provided a more anatomically representative solution than the previous 2D-based normalised DSM approach. SRRs were generated based on planned and accumulated FE-DSMs, indicating spatial clusters and localised dose patterns unique to each toxicity endpoint. These findings develop upon previous work investigating spatial aspects of rectal DSMs with regards to improving toxicity prediction [16], [27]. Several studies have used DSM analysis to demonstrate links between toxicity and mid-to-intermediate dose levels (i.e. other than maximum or prescription level doses) [5], [7], [10], [11], [12], [13], [15]. This suggests that further knowledge and understanding of the interaction between dose level, spatial distribution and location within the RW may improve endpoint-specific toxicity prediction.

From the SRR analysis presented here, of particular note is that although patterns differed per endpoint, regions of interest were generally identified away from the highest dose levels (i.e. closest to the prostate), and extended towards lower dose levels, posteriorly, superiorly and/or inferiorly, depending on the particular endpoint. This effect has been observed in previous studies, with several authors reporting associations between dose to the inferior rectum and rectal bleeding [8], [10], [12], [13], [18]. Dréan et al. [18] conducted voxel-level DVH analysis of planned dose to the rectal volume, and found rectal bleeding grade 1 to be associated with dose to a SRR in the inferior-anterior hemi-rectum. Onjukka et al. [10] evaluated planned dose to the combined RW and anal canal using DSMs, and found inferior regions corresponding to the anal canal to be predictive of faecal incontinence grade 1, and lateral-posterior subregions near the ano-rectal junction to be associated with rectal bleeding grade 2. Where endpoints are comparable, the SRRs determined in the present study generally, but not exclusively, encompass these regions of interest. The observed variation in SRR clusters and patterns supports the notion that discrete toxicity effects may have different pathophysiologies.

A novelty of the work presented here is not only the comparison of SRR and full RW dose-toxicity analysis for planned dose, but also for total accumulated dose. The rectum is highly susceptible to motion and deformation [31], and therefore planned dose is not equal to delivered dose [8], [9], [14], [16], [41]. When comparing planned and delivered dose, total accumulated daily dose to the RW was systematically lower than planned dose (p < 0.001). This effect has been reported previously [14], [16] and is thought to be due to interfraction motion causing a blurring of the highest dose regions. However, when focusing on SRRs, accumulated dose was greater than planned dose (p < 0.001), an effect masked when considering the full RW. This may indicate that these SRRs are more greatly affected by interfraction motion, and therefore dose calculated at treatment planning is not representative of the dose actually received in these regions. Differences between planned and delivered dose limit the level of accuracy achievable in dose-toxicity modelling.

In lieu of widely available systems for accumulating daily delivered dose to the rectum, previous studies have addressed this problem using approaches such as estimating motion-inclusive dose through statistical simulation [42] or by extrapolation of dose calculated using a sample of cone-beam CT (CBCT) scans acquired throughout the course of treatment [14], [41]. Casares-Magaz et al. conducted a case-control study of 24 patients whereby accumulated dose was estimated through rigid registration of dose from 13 CBCT scans per patient. Dose-surface histograms of RW DSMs were analysed for associations with gastrointestinal toxicity grade 2. Results revealed a subregion in the inferior rectum from the accumulated DSM predictive of toxicity, where planned dose could not be considered predictive. This supports the hypothesis that delivered dose is a better predictor of rectal toxicity than planned dose.

For accumulated dose, the use of SRRs produced stronger associations with toxicity than the full RW for three of four endpoints; rectal bleeding, proctitis, and faecal incontinence. For rectal bleeding this suggested radiosensitive regions spanning the inferior circumference to the right-posterior mid-RW, and an anterior-superior cluster. For proctitis, this was focussed in the posterior to lateral inferior RW. Small SRR clusters in the mid-posterior and right-inferior RW were identified for faecal incontinence. Results for stool frequency could not be considered significant, nor could any result from planned dose, other than EUD to the full RW for rectal bleeding. Of the three endpoints found to be associated with EUD to SRRs, although accumulated AUCs were greater than from planned dose, there was considerable overlap of CIs for proctitis and faecal incontinence, suggesting no significant advantage to using accumulated dose over planned dose. ORs from accumulated SRRs were greater than from planned SRRs for rectal bleeding and incontinence. All other results had CIs crossing 1. The greatest improvement to dose-toxicity association when analysing accumulated dose to SRRs was found for rectal bleeding, where the AUC was 13% greater than planned dose to the full RW and 15% greater than planned EUD to SRR, with a corresponding increase in OR of 26%.

The issue of correcting for multiple comparison testing in voxelwise dose-toxicity studies has been debated in the scientific literature. Palma et al. [43] promote correcting for family-wise error rate (such as Bonferroni correction) or false discovery rate (such as the Benjamini-Hochberg, as investigated here). However, others have suggested that p-value adjustment may lead to increasing sensitivity to the point of overfitting the data when applied to voxel-level DSM analysis [10]. In their analysis of bladder DSMs, Palorini et al. [44] discussed the issue at depth, and argued against reducing or adjusting p-values in this context, emphasising that their main goal was ’to identify the shape of dose regions that show the highest discrimination between the two groups of patients (toxicity versus no toxicity) without looking at significance at the single-pixel level’. They applied a similar post-processing smoothing approach to that used in this study. Based on these recommendations, Onjukka et al. [10] reported results for unadjusted p-values in a recent pattern analysis of rectal DSMs. Results for adjusted p-values were also included, but were considered overly restrictive. In the present study, results based on adjusted p-values have been included for completeness, but the focus was on SRRs generated using unadjusted p-values.

Higher AUCs were expected for accumulated dose than planned dose as, intuitively, true delivered dose incorporating interfraction motion should be more predictive of toxicity than the planned dose based on static anatomy. However, differences were generally small and there was overlap of CIs. The significance of results based on these small differences, for the relatively low toxicity rates, can therefore be highly sensitive to other limiting factors. For example, here, the accuracy of the FE model is dependent on the rectal contours, and accumulated dose data can be limited by the MVCT field of view.

The power of the current study was limited by its sample size. Although similar to previous studies [10], [18], here this resulted in relatively large confidence intervals, particularly when considering validation results with respect to training. The number of validation data points was lower than that generally considered adequate for ROC/AUC analysis, so OR comparisons were also included to provide further interpretation of results. Analysis of the validation set could not be considered to verify the training set results. However, these results should be interpreted with caution. The poor performance and wide confidence intervals are likely due to the small sample size and event rate within the validation set. The rationale for using a data partitioning approach (rather than alternative internal validation methods such as bootstrapping or k-fold cross-validation) was to ensure that the testing of the SRR analysis was conducted using an independent dataset, kept blind from the SRR-generation process. This was the approach used by Dréan et al. [18] in their work determining rectal SRRs using a similar sample size.

The assumption was made that the two fractionation regimes could be radiobiologically combined due to equivalent levels of toxicity incidence [45]. This approach has been used in similar investigations of rectal toxicity [10], [18]. However, within the patient cohort, increased rates of faecal incontinence and stool frequency were observed in the 74 Gy arm. Conversely, in a recent study by Heemsbergen et al. [15], increased rates of toxicity were observed for rectal bleeding and stool frequency in a hypofractionated arm (64 Gy in 3.4 Gy fractions) compared with the conventional fractionation (78 Gy in 2 Gy fractions). The current study does not consider patient co-factors (including prescription), but such effects may be investigated in the future when a larger patient cohort is available.

The incorporation of accumulated dose to SRRs into dose-toxicity analysis has been shown to be more beneficial than using planned dose alone for certain toxicity endpoints, particularly rectal bleeding. The inclusion of accumulated dose into toxicity prediction models becomes an increasingly important consideration with the move towards hypo and ultra-hypofractionated treatments such as prostate SABR, where daily interfraction motion has a much greater effect in terms of contribution towards total delivered dose. Additionally, the duration of these treatment deliveries is of the order of magnitude where intrafraction motion should also be considered [46]. The effect of intrafraction motion was outwith the scope of the VoxTox research programme.

In conclusion, we have demonstrated that voxel-level dose accumulation using FE-DSMs facilitates higher resolution interrogation of associations between rectal SRRs and toxicity. Regions of significant dose differences were identified per endpoint, away from the highest dose levels. SRRs defined from accumulated dose were more predictive of rectal bleeding than from planned dose. However, validation results were not significant. Ultimately, the aim is to use voxel-level accumulated dose information to develop more accurate and robust toxicity prediction models than previously achievable from planned dose. In the context of adaptive radiotherapy, these models could be used to identify and monitor patients most at risk of developing toxicity. If required, the treatment could then be modified to minimise and control the risk of toxicity, with a view to reducing incidence rates, lowering medical costs of associated care, and improving the patient’s quality of life.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.