Investigating rectal toxicity associated dosimetric features with deformable accumulated rectal surface dose maps for cervical cancer radiotherapy

The combination of the external beam radiotherapy (EBRT) and brachytherapy (BT) (EBRT+BT) is a common therapy regime for locally advanced cervical cancer [1]. Recent monocentric and multicentric EBRT+BT studies [2‐4] have shown promising results with high tumor local control rate. However, radiation induced side effect (or toxicity) on organs at risk (OARs), such as rectum, bladder and vagina, is still a concern. Serious side effects such as bowel obstruction can occur months to years after treatment and impact negatively on the patients’ quality-of-life. The correlation between OARs’ morbidity and radiation dose parameters was analyzed in EMBRACE study [2]. Particularly, the D_0.1cm³, D_1cc and D_2cm³ of rectum were used to establish dose-toxicity relationship in the occurrence of rectal morbidity. D_0.1/1/2cm³ are conventional dose volume parameters (DVPs) extracted from dose volume histogram (DVH). Inherently, they are in deficiency of dosimetric spatial information. Studies have shown close relationship between the spatial dose characteristics and rectal toxicity [5‐12]. For instance, Wortel et al. [12] observed significant differences in local rectal dose distribution between prostate cancer patients with and without toxicity by utilizing the unfolded two dimensional (2D) rectum surface dose map (RSDM). Similarly, Munbodh et al. [10] demonstrated that late rectal toxicity was related to dose on the upper rectum region by investigating dose pattern on the RSDM. Buettner et al. [6] analyzed the RSDM and found significant correlation between the subjective sphincter control and the dose delivered to the anal sphincter region. Another issue with current D_0.1/1/2cm³ evaluation procedure is that the cumulative dose is summed with an assumption that the hotspot regions are stationary throughout the entire fractional treatments [13, 14]. However, this static assumption is often violated by the large inter-fraction rectum deformation, especially in intra-cavity brachytherapy treatment cases [15‐17]. Recently, promising advancements have been reported by Moulton et al. who investigated the associations between RSDM and gastrointestinal toxicities after deformably registering each phase of a combined EBRT-BT prostate cancer treatment [18]. These limited but inspiring studies shed light on the possibility of revealing more accurate dose-toxicity relationship by exploring the spatial dose distribution patterns on the deformable accumulated dose.

In this study, we proposed and evaluated a rectum dose-toxicity prediction scheme using both dose volume parameters and dose map spatial information. In addition, the accumulated rectal dose maps are obtained with the aid of an accurate deformable image registration. The accumulated 3D rectal surface dose was flattened to obtain a 2D RSDM. The DVPs were extracted from the DVHs of cumulative dose, while the texture features and the dose geometric parameters (DGPs) were extracted from the 2D RSDM. Representative features further computed from DVPs, textures and DGPs by principle component analysis (PCA) and statistical analysis were respectively feed into a support vector machine (SVM) equipped with a sequential feature selection (SFS) procedure. The predictive powers of the representative features were compared with the GEC-ESTRO dosimetric parameters D_0.1/1/2cc.

An effective rectal toxicity prediction scheme is essential for guiding radiation treatment planning. D_0.1/1/2cm³ are recommended by the GEC-ESTRO guidelines [36] for rectum dose monitoring, however, their predictive capabilities for rectal toxicity are still under investigation. Other studies reported that the D_5cc may be a more reliable estimate than other dose volume parameters to predict risks of rectosigmoid mucosal changes and late rectal complications [37, 38]. All these studies essentially used DVPs to predict rectal toxicity. The findings in current work align with the previous studies, e.g., statistically significant differences were observed in DVPs ranging between D_0.1cm³~D_6cc. However, better prediction was accomplished by utilizing all the significant DVPs in F_sta when compared with merely using the SA-D_0.1/1/2cm³ (Table 1).

In this study, we have compared two approaches, i.e., the PCA analysis and the statistical analysis, in extracting representative features for feeding the SVM-SFS prediction scheme. The merit of performing PCA is to reduce the number and correlation of the potential features by converting the features into a set of values of linearly uncorrelated variables (F_PCA in this study). However, these converted values in the PCA domain carry no physical meanings, it is therefore difficult to interpret the prediction why the F_PCA features are responsible for yielding corresponding prediction result. On the other hand, the statistical analysis approach reserves the physical meanings of the features by statistically pre-screening the significant features (F_sta in this study), and the comparison evaluations also showed superior performance if all the features in F_sta were used for prediction, when compared with F_PCA.

Recently, researchers started to investigate the prediction model with spatial dose information. For instance, Buettner et al. presented a late rectal toxicity method based on the parameterized representation of the 3D rectal dose [39]. Lee et al. proposed a metric based on both surface dose and distance to predict incidence of the rectal bleeding in prostate cancer patients treated with radical radiotherapy [40]. Drean et al. identified rectal subregions at risk of rectal bleeding by performing voxel-wise analysis on the rectal dose distribution [7]. In this study, we took advantage of the hollow structure of the rectum and flattened the 3D rectal dose to 2D RSDM to establish dose map toxicity prediction scheme. Though the RSDM neglects the doses in the rectum thickness direction, it preserves spatial dose information. The texture features and the DGPs, which are crafted to capture spatial dose distribution characteristics from the 2D RSDM, are able to provide more geometric and positional dosimetric information. Pioneer studies have shown potential correlations of spatial dose characteristics with rectal toxicity. For example, Drean et al. reported that the rectal subregions at risk of rectal bleeding are primarily located in the subprostatic anterior hemi-rectum and upper part of the anal canal [7]. Kim et al. found substantial correlation between rectal toxicity and percent rectal circumference at certain dose levels. Similarly, in this study, we have seen texture features and the geometric dosimetric features had better predictive power than the DVPs. These results hint us that rectum’s response to dose might be dose-spatial dependent. As shown in Table 2 and Fig. 4, geometric feature such as the relative area, the perimeter and the length were found to be associated with rectal toxicity. The toxicity group tended to have larger dose coverage on the high dose region (64, 67 and 68Gy). This finding was in agreement with previous studies that rectal bleeding was significantly correlated with high-dose metrics [25, 41‐43]. In addition, although only four DGPs were investigated in this work, other DGPs which were explored in previous studies also indicated associations with rectal toxicity. For instance, Buettner et al. investigated the eccentricity of the fitted ellipse of the dose region and found associations of the eccentricity with loose stools [5, 39, 44]. Moulton et al. also reported that compactness, circularity and confinement to the ellipse fits were correlated with rectal bleeding [18]. Incorporating these spatial features, which are crafted to depict the dose coverage and the shape of dose distribution, into the current model may potentially improve the predictive performance. Besides, two texture features, i.e., the HGZE and Complexity, were statistically different between the two groups, where the toxicity group has lower HGZE value (the Median(IQR) 317.81(51.19) vs. 376.63(37.70), p = 0.0007) but higher Complexity value (the Median(IQR) 130.71(9.04) vs. 119.77(10.18), p = 0.0087). However, how these texture features impact on rectal toxicity is still unclear. We applied the texture features on the RSDM with the intention to describe localized dosimetric patterns on the RSDM which are usually difficult to be noticed by human eyes. Yet, the drawbacks of the texture features are their deficiency of physical implications on interpreting correlations. Moreover, feature stability, e.g., whether the selected texture feature will change if different patient cohort size are used, is still an exploratory issue [45‐47]. Since current work is a pioneer feasibility study of applying texture feature analysis on the deformably accumulative rectum surface dose map, more in-depth investigations on a larger patient cohort is still required in the future.

In this study, the generated 2D RSDM reserves the physical length of the rectum in both the superior-inferior direction and the circumferential direction on each slice (see Additional file 1: Appendix B). This was to ensure that the geometric features (e.g., area, perimeter, length, etc.) extracted from the RSDM would carry physical implications to signify the scale of dose delivered on rectum surface. Note that the drawback for reserving the physical dimension of the rectum is that the inter-patient variations of the rectum size could possibly influence/mask the significances of differences of the extracted geometric features between the two groups, especially given a small patient cohort. But the rectum size tends to be a random number across patients and therefore its impact would decrease and be minor in a larger patient cohort.

In addition, reporting accurate accumulated dose over the entire treatment course is a nontrivial task because of the substantial inter-fractional rectum deformation exists in the BT treatments. In this work, a previously developed TOP-DIR algorithm was used, although accurate geometric registration accuracy had been achieved and validated on a porcine bladder phantom (~ 2 mm), further phantom studies are still needed to justify its effectiveness in rectum registration, and the dosimetric errors in the subsequent dose summation step also need to be monitored.

In this study, we added the EBRT dose to the accumulated BT dose without deformation. The reasons are twofold: firstly, a homogenous dose distribution (hot spot < 107%) often covers the entire pelvic region in our EBRT treatment plan regimen. Often large portion of rectum are within treatment fields, especially for 3D plans. Only a very small inferior portion of the rectum is outside of the large pelvis treatment fields and dose variation across rectum is often within 15%. With this relative homogenous dose in a large region across the pelvis, we could assume rectum receiving a consistent and homogenous dose in EBRT as planned. In this study, the EBRT plans were generated with 4-field box 3D plans or 9 field IMRT plans. Theoretically, these two techniques on a same patient would result in different EBRT dose distributions due to dose conformity and hence different accumulated dose on RSDM. It is therefore more appropriate to investigate the extracted features for each technique. However, it is impractical to implement in current study since only a small patient sample was available. The influence of these two EBRT techniques on the stability of the extracted features still needs further investigations on a larger patient cohort.

On the other hand, DIR between EBRT and BT CTs is challenging because of the clinical use of the intracavitary applicator in BT. Registering the BT CT image with applicator to the EBRT CT image without applicator (or vice versa) is difficult, if not impossible, since the point-to-point correspondence assumption is usually violated in most DIR algorithms. Consequently, the dosimetric uncertainties via EBRT-BT DIR might be possibly even larger than that summed without deformation. There are several reported attempts to address this issue [48, 49], for example, Berendsen et al. [48] proposed a DIR with penalty term that minimizes the volume of the missing structure for cervical MR images with and without applicator. Vasquez Osorio et al. [49] validated a structure-wise registration with vector field integration to map the largely deformed anatomies between EBRT and BT. However, the EBRT-BT DIR needs to be treated prudently, and these novel methods need comprehensive validations before they can be confidently applied in a clinical setting. Adding EBRT to BT without deformation is therefore a reasonable approximation without knowing the uncertainties brought by the EBRT-BT DIR.

The choice of prediction models and feature selection strategies may also affect the predictive performance. We used the SVM-SFS scheme because it is the most common method to construct a predictive model with simultaneously feature selection. Though satisfactory performances have been achieved, other predictive models (e.g., random forest classifier) or feature selection methods (e.g., clonal selection algorithm) can provide even better predictive accuracy [50, 51].

For screening of the representative feature F_sta, the unadjusted p-values were used for statistical analysis, however, the current findings will probably change if the p-values were corrected for multiple testing. In fact, p-value adjustment is restrictive to application with many tests and applying it in the context of RSDM analysis is still controversial [18, 52]. Since the physical length of the rectum was reserved on the RSDM in this study, the resolutions of the RSDMs were essentially patient specific. Multiple testing might not be applicable for this scenario where the resolution of the RSDM is fixed for each patient. Even though p-values corrections have been reported in other similar investigations using RSDM for rectal toxicity studies, however, the adjusted p-values did not demonstrate clear trends across regions on the RSDM where only limited and isolated regions of significance were found after applying multiple testing correction in RSDM analysis [18, 52, 53]. Furthermore, reporting the raw p-values is an exploratory study of finding predictive factors correlated to rectal toxicity without the risk of missing important factors which might be discarded if found to be insignificant after p-value adjustment [18].

One limitation of current study is that the patient cohort was small. Thus, the number of extracted features was larger than the patient sample size. To reduce the chance of getting over-fitting, a statistical analysis was performed to screen out the significant features before feeding into the predictive model. This guarantees a more robust feature ranking in the subsequent feature selection step in SVM-SFS. But note that a more effective way to observe overfitting is to separate the patient cohort into three datasets, i.e., one for training, one for validation and hyper-parameter tuning and one for testing. However, it was impossible to effectively separate our samples into three datasets, and cross validation was therefore our secondary option for model performance observation in a small patient cohort. But overfitting might also occur in the cross-validation space attributed to other factors such as the quantity of features considered, the selection of model hyper-parameters, etc., therefore, larger patient data is a key for evaluating model stability and generalization capability.

Another limitation of the study is that our study is purely on dosimetric parameter without consider clinical factors. Multivariable modeling of radiotherapy outcomes has been conducted by El Napa et al. [54]. We will further include clinical factors in our near future studies.

In summary, we have proposed and validated a rectum toxicity prediction method based on an accurate point registration and machine learning for cervical cancer radiotherapy. The extensive experimental results have demonstrated the feasibility of the proposed scheme for rectal toxicity prediction, rendering it a potential tool for clinical OARs dose control and complication prediction.