STructural Rectal Atlas Deformation (StRAD) Features for Characterizing Intra- and Peri-wall Chemoradiation Response on MRI

Abstract

Radiomic features which quantify morphologic texture and shape of tumor regions on imaging have found wide success in characterizing treatment response in vivo. A more detailed interrogation of intra- and peri-tumoral regions for response-related cues could be achieved by capturing subtle structural deformations that occur due to tumor shrinkage or growth. In this work, we present a novel suite of STructural Rectal Atlas Deformation (StRAD) features to quantify tumor-related deformations in rectal cancers via a cohort of 139 patient MRIs. In flexible non-rigid organs such as the rectum, inter-patient differences complicate evaluation of tumor-related deformations that may occur within the rectal wall or in the peri-rectal environment; necessitating construction of a canonical rectal imaging atlas. Using 63 pelvic MRIs where healthy rectums could be clearly visualized, we built the first structural atlas for the healthy rectal wall. This atlas was used to compute structural deformations within and around locations in the rectal wall of patients where tumor was present, resulting in intra- and peri-wall StRAD descriptors. We evaluated the efficacy of our StRAD features in 2 different tasks: (a) predicting which rectal tumors will or will not respond to therapy via baseline MRIs (n = 42), and (b) identifying which rectal tumors were exhibiting regression on post-chemoradiation MRIs (n = 34). Using a linear discriminant analysis classifier in a three-fold cross-validation scheme, we found that intra-wall deformations were significantly lower for responders to chemoradiation; both on baseline MRIs (AUC = $0.73\pm 0.05$) as well as on post-therapy MRIs (AUC = $0.87\pm 0.03$). By comparison, radiomic texture features for both intra- and peri-wall locations yielded significantly worse classification performance in both tasks.

J. Antunes and Z. Wei—Joint first authors.

Research supported by NCI (U24CA199374-01, R01CA202752-01A1, R01CA208236-01A1, R01CA216579-01A1, R01CA220581-01A1, U01CA239055-01, F31CA216935-01A1), NCRR (1C06RR12463-01), DOD/CDMRP (W81XWH-15-1-0558, W81XWH-16-10329, W81XWH-18-0404, W81XWH-18-1-0440), NIDDK (1P30DK09794 Pilot Award), NBIB (T32EB00750912), VA (IBX004121A Merit Review Award), the Dana Foundation David Mahoney Neuroimaging Program, the Ohio Third Frontier Technology Validation Fund, the Wallace H. Coulter Foundation Program in the Department of Biomedical Engineering and the Clinical and Translational Science Award Program at CWRU. Content solely responsibility of the authors and does not necessarily represent the official views of the NIH, USDVA, DOD, or the United States Government.

1 Introduction

The advent of radiomics has demonstrated great success for predicting and evaluating treatment response via imaging in different cancers [1]. Radiomic approaches have typically extracted morphologic texture or shape descriptors of the tumor region, which have been related to underlying pathologic or molecular heterogeneity that drive therapy response [2]. As an example, prediction of response to chemoradiation in rectal cancers via pre- or post-treatment MRI [3, 4] has been limited to morphologic radiomic descriptors for image appearance alone. Unlike deep learning approaches (which are data-driven solutions to lesion segmentation, localization, or detection [5]), radiomics also leverages “hand-crafted” descriptors to quantify specific imaging characteristics both within and around the tumor [6]. For instance, new classes of features that quantify tissue deformations or surface distentions on imaging have been linked to aggressive tumor growth [7] as well as tumor recurrence [8] (based on available reference atlas representations in solid organs such as the brain or the prostate). Quantifying such structural changes in more flexible organs such as the rectum requires construction of a healthy rectal wall atlas (i.e. rectal anatomy without tumor). We hypothesized that constructing a healthy rectal atlas could allow for unique quantification of disease-specific structural changes in the rectal environment (wall/tumor, peri-wall/tumor) that may be closely related to tumor response to therapy. The novel contributions of this work are:

• Development of the first structural atlas representation for healthy rectal wall anatomy, via a multi-stage registration scheme using pelvic MRIs (from other cancers) where normal rectums are visible.

• The first attempt at relating subtle structural deformations occurring within and around rectal wall regions against chemoradiation-related tumor growth or shrinkage in vivo.

Our novel STructural Rectal Atlas Deformation (StRAD) features were evaluated in the context of 2 distinct clinical problems in rectal cancer: (a) prediction of pathologic non-responders to chemoradiation via baseline treatment-naïve MRI, and (b) assessment of pathologic responders on post-chemoradiation MRI. Together, these represent the major clinical challenges facing personalization of patient management in rectal cancers.

Fig. 1.

Overall workflow for extracting StRAD descriptors including construction of healthy rectal atlas, computing deformations within and around the rectal wall in rectal cancer patients (with respect to the atlas), and extracting structural deformation features for intra- and peri-wall regions.

2 Methodology

Quantifying structural deformations within and around the rectum involves 3 major steps: (i) building a structural atlas for normal rectal wall anatomy on imaging [9], (ii) computing structural deformations of the rectal wall in patients with tumors with respect to this atlas [10], and (iii) extracting tumor-related structural deformation descriptors within the rectal wall and peri-rectal environment. The steps to extract StRAD features are illustrated in Fig. 1.

2.1 Construction of Structural Rectal Atlas

A set of N MRI scenes depicting the healthy rectum is utilized, denoted $\mathcal {X} = (C,f)$, where C is a 3-dimensional spatial grid and f(c) represents the MRI intensity at each voxel $c \in {C}$. The primary anatomic region defined within this MRI scene is the healthy rectal wall, denoted $\mathcal {X}^{r} = (C,f^r)$, where $f^r(c) = 1$ within the rectal wall and zero in the rest of the scene. $\mathcal {X}^r$ can be identified and annotated by experts on all $\mathcal {X}$, and is depicted via the color green in Fig. 1.

The final output of this step is the healthy rectal wall atlas, denoted $\mathcal {A}=(C,g^r)$. We define $g^r(c) \in [0, 1]$, as the frequency of a particular location $c\in C$ where $f^r(c)=1$ (i.e. corresponding to rectal wall); across N different input subject scenes. These N different subject scenes are aligned to a registration template for projection into a canonical space to construct $\mathcal {A}$, via the following 3 transformations [9]:

• 1. A simple transformation, $\tau _\rho$, is used to map N different subject scenes $\mathcal {X}$ such that they are all centered and isotropically scaled in X, Y, and Z axes. The resulting initial atlas, $\mathcal {A}_\rho$, is therefore not dependent on selecting a specific subject as the template and is constructed such that $A_\rho =(C,g^r$), where $g^r(c) = \frac{1}{N}\sum _N f^r(c)$, for every location $c \in C$, across all N studies after $\tau _\rho$ has been applied (i.e. $g^r(c)$ is the frequency of a location corresponding to the rectal wall).

• 2. Affine registration is used to compute $\tau _\alpha$ for projecting all $\mathcal {X}$ onto $\mathcal {A}_\rho$. The affinely transformed subject scenes are used to construct $A_\alpha =(C,g^r$) (based on re-computing $g^r(c)$ $\forall c\in C$, across all N studies).

• 3. Deformable registration is used to align $\mathcal {X}$ to $\mathcal {A}_\alpha$. The final structural rectal atlas $A=(C,g^r$), is constructed based on re-computing $g^r(c)$ $\forall c\in C$, across N deformed subject scenes.

2.2 Computing Structural Deformations with Respect to the Atlas

Given a patient MRI scene, denoted $\mathcal {I}$, structural deformations in the rectal environment are quantified [10] with respect to the healthy atlas $\mathcal {A}$. The rectal wall within the patient MRI scene is denoted $\mathcal {I}^r$. First, $\mathcal {A}$ is non-rigidly registered to $\mathcal {I}$ using a normalized mutual information-based similarity measure within a b-spline registration scheme. This non-rigid alignment is formulated as ($\mathcal {I}^r,\mathcal {I})=T(\mathcal {A})$, where T is the forward transformation of the composite voxel-wise deformation field (comprising affine and deformable components) that maps the rectal wall between the reference ($\mathcal {I}^r$) and floating ($\mathcal {A}$) volumes. This transformation is then inverted to yield $T^{-1}$, which is used to map $\mathcal {I}$ into the $\mathcal {A}$ space. This 2-stage mapping process is designed to compute structural deformations within $\mathcal {I}$ with respect to $\mathcal {A}$ at every $c\in C$, hypothesized to occur as a result of tumor-related growth or shrinkage of the rectal wall.

2.3 Extracting StRAD Descriptors for Subregions Within Rectal Wall and Peri-rectal Environment

Structural deformations are quantified for each rectal cancer patient scene within $\mathcal {I}^r$, as well for a peri-wall area denoted $\mathcal {I}^p$. The latter was defined based on $\mathcal {I}^r$ within each of the experiments later conducted. Once $\mathcal {I}$ is mapped to the $\mathcal {A}$ space, all voxel positions $(c_x,c_y,c_z)$ are assumed to be displaced by $[\delta x,\delta y,\delta z]$, to result in $(c_x',c_y',c_z')=(c_x,c_y,c_z)+(\delta x,\delta y,\delta z)$. Based on this displacement vector, the deformation magnitude is computed as $D(c)=\root \of {(\delta x)^{2}+(\delta y)^{2}+(\delta z)^{2}}$, for every $c \in C$. The descriptor $\mathbb {F}_{def}^r$ for intra-wall deformations comprises first order statistics (i.e. mean, median, standard deviation, skewness, and kurtosis) of D(c) for all the voxels c within the rectal wall $\mathcal {I}^r$. Similarly, the peri-wall deformation descriptor $\mathbb {F}_{def}^p$ can be computed based on first-order statistics of the deformation magnitudes in $\mathcal {I}^p$.

3 Experimental Design

3.1 Data Description

Healthy Rectum Cohort ($S_1$): A cohort of 63 patients who had been diagnosed with prostate cancer and had undergone an axial pelvic MRI scan were curated. As no endorectal coil had been used during the MRI scans, these images provided a clear in vivo visualization of the healthy rectal wall.

Baseline RCa Cohort ($S_2$): A cohort of 42 patients who had been diagnosed with rectal cancer were identified, all of whom had undergone axial 3 Tesla (T) T2w MR imaging before standard-of-care chemoradiation. The goal was to predict non-responders to chemoradiation using this baseline MRI scan. Pathologic tumor stage (T-stage, based on excised rectal specimens) was used as a marker of response, where ypT3-4 corresponded to extensive tumor being present in the specimen despite chemoradiation. Based on this pathologic classification, $n=22$ patients were identified as being non-responsive to chemoradiation (ypT3-4), and the remainder as responders to chemoradiation (ypT0-2, $n=20$).

Post-therapy RCa Cohort ($S_3$): A separate cohort of 34 RCa patients was curated, where patients had axial 3 T T2w MRIs available after undergoing standard-of-care chemoradiation but prior to excision surgery. In this cohort, the goal was to identify which patients exhibited marked tumor regression (based on pathologic T-stage) via the post-therapy MRI scan. With ypT0-2 indicating minimal or dying tumor within the rectal wall after chemoradiation, $n=17$ patients were assessed as being responders and the remaining $n=17$ were classified as exhibiting minimal or no response to chemoradiation (ypT3-4).

3.2 Implementation Details

For all 139 MRI scans in cohorts $S_{1-3}$, the entire length of the visible rectal wall from the anus to the peritoneal reflection was annotated by an expert radiologist. For the 76 RCa cases in $S_2$ and $S_3$, the slices most suspicious for tumor presence were also identified by the radiologist (using anatomic information from pathology reports). The healthy atlas $\mathcal {A}$ was constructed using $N=63$ MRI pelvic scans in $S_1$ using the approach from Sect. 2.1. Evaluation of the atlas in terms of overlap in annotated rectal wall as well as internal lumen regions (across all patients in $S_1$ after deformable mapping) yielded a Dice similarity coefficient of 0.87, indicating $\mathcal {A}$ was a relatively accurate representation.

Deformation fields for the remaining 79 RCa scans in $S_2$ and $S_3$ (with respect to $\mathcal {A}$) were then computed to yield intra-wall and peri-wall StRAD descriptors, $\mathbb {F}_{def}^r$ and $\mathbb {F}_{def}^p$ respectively (each a $5 \times 1$ vector). The peri-wall region was empirically defined as an 8 pixel band along the outer wall boundary for $S_2$ and $S_3$. All registration steps were implemented using elastix [11], with a grid spacing of $9\times 9\times 9$ when computing b-spline deformations. Radiomic texture features were also extracted to characterize the appearance of intra- and peri-wall areas on all 79 RCa scans [1], yielding $\mathbb {F}_{tex}^r$ and $\mathbb {F}_{tex}^p$ (each a $825\times 1$ vector). Both deformation and texture features were extracted from 3 consecutive slices comprising the largest wall area suspicious for tumor, assuming that this region was most likely to exhibit signatures related to tumor growth or shrinkage on MRI.

Separate experiments were conducted using each of $S_2$ and $S_3$ in a cross-validation setting, with the goal of distinguishing between the 2 patient groups in each cohort. Following feature extraction, minimum redundancy maximum relevance feature selection [12] (mRMR) was used to identify the 3 most relevant features within each of $\mathbb {F}_{def}^r$, $\mathbb {F}_{def}^p$, $\mathbb {F}_{tex}^r$, and $\mathbb {F}_{tex}^p$. The most relevant set of features from each vector was then evaluated via a Linear Discriminant Analysis (LDA) classifier. A total of 50 iterations of a three-fold (one fold held-out for testing), patient-stratified, cross-validation scheme were utilized to ensure robustness of feature selection and classifier evaluation steps; with ROC analysis for evaluation. These steps were repeated for each of $S_2$ and $S_3$, and the area under the ROC curve (AUC) across all cross-validation runs was used to compare each of $\mathbb {F}_{def}^r$, $\mathbb {F}_{def}^p$, $\mathbb {F}_{tex}^r$, and $\mathbb {F}_{tex}^p$ (via Wilcoxon ranksum testing) to determine which feature set was most relevant for treatment response characterization.

Fig. 2.

(a), (c) Representative baseline T2w MRI scans from $S_2$ for two different patients showing deformation field as colored arrows within the rectal wall (annotated in yellow). (b), (d) Corresponding intra-wall deformation magnitudes visualized as a heatmap, where red corresponds to higher D(c). (e) Boxplots of skewness in deformation magnitudes reveal intra-wall deformations in non-responders to chemoradiation are positively skewed (i.e. larger magnitudes in NR patients, ypT3-4), compared to responders (GR). (f) Bar plot of AUC values for different feature descriptors, where $\mathbb {F}^r_{def}$ (blue) resulted in a significantly higher performance than $\mathbb {F}^p_{def}$ (red), $\mathbb {F}^r_{tex}$ (ochre), and $\mathbb {F}^p_{tex}$ (purple). (Color figure online)

Fig. 3.

Representative post-therapy T2w MRI scans from $S_3$ for two different patients showing deformation field (visualized as colored arrows) within yellow outline of rectal wall. (b), (d) Corresponding intra-wall deformation magnitudes for these patients visualized as heatmaps (red corresponds to higher D(c)). (e) Boxplots of standard deviation of deformation magnitudes within the rectal wall reveal significantly less variable deformations associated with responders to chemoradiation (GR, ypT0-2), compared to non-responders (NR). (f) Barplot of AUCs for different feature descriptors, where $\mathbb {F}^r_{def}$ (blue) resulted in a significantly higher performance than $\mathbb {F}^p_{def}$ (red), $\mathbb {F}^r_{tex}$ (ochre), and $\mathbb {F}^p_{tex}$ (purple). (Color figure online)

4 Results and Discussion

4.1 Experiment 1: Predicting Non-responders to Chemoradiation via Baseline MRIs

The most relevant StRAD descriptors identified in experimental evaluation of $S_2$ were the skewness and standard deviation of intra-wall deformation magnitudes, the former of which is visualized in Fig. 2. Higher deformation magnitudes are shown in red in Fig. 2(b) and (d), while lower magnitudes are in blue. Our results indicate that non-responders to chemoradiation may be associated with significantly higher structural deformations within the rectal wall on baseline MRI scans (Fig. 2(e), positive skew associated with non-responders), when compared to the healthy rectal atlas. This resonates with previous findings where it has been reported that smaller rectal tumors tend to respond favorably to chemoradiation [13], which would result their being associated with less pronounced wall deformations (with reference to a healthy atlas). Further, the intra-wall StRAD descriptor ($\mathbb {F}^r_{def}$) also yielded the best overall AUC in this classification task (blue bar in Fig. 2(f), $0.73\pm 0.05$). This was significantly higher ($p<0.001$) than the AUCs for each of $\mathbb {F}^p_{def}$, $\mathbb {F}^r_{tex}$, and $\mathbb {F}^p_{tex}$.

4.2 Experiment 2: Identifying Responders After Chemoradiation via Post-therapy MRIs

In $S_3$, the median and standard deviation of the intra-wall deformation magnitude were identified as the most relevant StRAD descriptors. Figure 3 visualizes representative heatmaps of intra-wall deformation magnitudes on post-therapy MRIs, indicating that responders are associated with significantly lower and less variable structural intra-wall deformations (Fig. 3(e), blue corresponds to lower magnitude in heatmaps). As non-responders (i.e. ypT3-4) are likely to have more tumor extent outside the rectal wall despite chemoradiation [14], this would be reflected in the rectal wall being more deformed with respect to the healthy rectal atlas. The intra-wall StRAD descriptor ($\mathbb {F}^r_{def}$) significantly outperformed all of $\mathbb {F}^p_{def}$, $\mathbb {F}^r_{tex}$, and $\mathbb {F}^p_{tex}$ in terms of AUC values for this classification task ($0.87\pm 0.03$, $p<0.001$, note blue bar in Fig. 3(f)).

5 Concluding Remarks

In this study, we presented novel STructural Rectal Atlas Deformation (StRAD) features for characterizing intra- and peri-wall response to chemoradiation on rectal MRIs. Our study involved construction of a reference healthy rectal wall atlas, which was then applied to compute tumor-related deformations on baseline and post-chemoradiation MRIs separately. StRAD features from within the rectal wall were found to be most effective for characterizing tumor treatment response on MRI. Non-responder RCa patients in both pre- and post-therapy settings were found to be associated with significantly higher and more variable intra-wall deformations; likely occurring as a result of more aggressive tumor growth. By contrast, morphologic texture features performed significantly worse both for predicting as well as evaluating response to therapy via MRI. Future work will include validation of StRAD features on a larger, multi-site cohort of data as well as evaluation of parameter sensitivity. Additionally, we will integrate StRAD features with morphologic descriptors and clinical variables to reliably predict and assess treatment response for rectal cancers in vivo.