Registration-based biomarkers for neoadjuvant treatment response of pancreatic cancer via longitudinal image registration

Abstract.

Purpose

Pancreatic ductal adenocarcinoma (PDAC) frequently presents as hypo- or iso-dense masses with poor contrast delineation from surrounding parenchyma, which decreases reproducibility of manual dimensional measurements obtained during conventional radiographic assessment of treatment response. Longitudinal registration between pre- and post-treatment images may produce imaging biomarkers that more reliably quantify treatment response across serial imaging.

Approach

Thirty patients who prospectively underwent a neoadjuvant chemotherapy regimen as part of a clinical trial were retrospectively analyzed in this study. Two image registration methods were applied to quantitatively assess longitudinal changes in tumor volume and tumor burden across the neoadjuvant treatment interval. Longitudinal registration errors of the pancreas were characterized, and registration-based treatment response measures were correlated to overall survival (OS) and recurrence-free survival (RFS) outcomes over 5-year follow-up. Corresponding biomarker assessments via manual tumor segmentation, the standardized response evaluation criteria in solid tumors (RECIST), and pathological examination of post-resection tissue samples were analyzed as clinical comparators.

Results

Average target registration errors were $2.56\pm 2.45\mathrm{mm}$ for a biomechanical image registration algorithm and $4.15\pm 3.63\mathrm{mm}$ for a diffeomorphic intensity-based algorithm, corresponding to 1–2 times voxel resolution. Cox proportional hazards analysis showed that registration-derived changes in tumor burden were significant predictors of OS and RFS, while none of the alternative comparators, including manual tumor segmentation, RECIST, or pathological variables were associated with consequential hazard ratios. Additional ROC analysis at 1-, 2-, 3-, and 5-year follow-up revealed that registration-derived changes in tumor burden between pre- and post-treatment imaging were better long-term predictors for OS and RFS than the clinical comparators.

Conclusions

Volumetric changes measured by longitudinal deformable image registration may yield imaging biomarkers to discriminate neoadjuvant treatment response in ill-defined tumors characteristic of PDAC. Registration-based biomarkers may help to overcome visual limits of radiographic evaluation to improve clinical outcome prediction and inform treatment selection.

1. Introduction

Pancreatic ductal adenocarcinoma (PDAC) represents the most prevalent primary cancer of the pancreas and is associated with 5-year survival rate of only 10%, among the lowest for common primary cancers of any organ.¹ Due to late onset of symptoms, PDAC is often diagnosed at advanced stages where fewer than 20% of patients are initially eligible for surgical resection.² Due to poor prognosis, neoadjuvant chemotherapy may be administered to control the risk of local and metastatic progression, to screen resection candidates for highly aggressive disease, and less commonly to downstage unresectable PDAC.³ Consequently, careful radiographic observation is necessary to evaluate treatment response. However, PDAC lesions often present with ill-defined boundaries and subtle contrast enhancement patterns that make these tumors extremely difficult to precisely delineate in computed tomography (CT) and magnetic resonance imaging, even for highly skilled radiologists.⁴ While segmentation of the outer pancreas is less variable, multiple studies have demonstrated that the inter-observer variability of manual segmentation of PDAC tumors is quite high, with reported Dice overlaps of merely 0.7 in both modalities.^5–7 These inconsistencies arise due to shallow contrast gradients that obscure a precise boundary for the lesion relative to surrounding parenchyma, and variability is likely to further increase in post-treatment imaging due to fibrosis of viable tumor.

This inability to accurately annotate lesion morphology over time ultimately confounds conventional radiographic evaluation of treatment response, impacting both the intra-reader consistency of tumor diameters measured for radiological reports and subsequent longitudinal assessments that may rely on reports generated by different readers. These factors have contributed to key studies suggesting that the current clinical standard of the response evaluation criteria in solid tumors (RECIST) may not adequately predict PDAC treatment outcomes following neoadjuvant therapy.⁸ Yet, other retrospective studies dispute these findings and suggest that imaging measures of PDAC tumor volume indeed correlate with clinical outcomes.^9,10 Although changes in tumor size after neoadjuvant chemotherapy are likely to be clinically relevant response variables, poor reliability of lesion measurement throughout the PDAC treatment course limits integration of these non-invasive imaging markers with the clinical decision-making paradigm. To more precisely assess treatment response, this study proposes a registration-based approach to longitudinally track and non-invasively quantify changes in tumor size throughout the treatment course. This registration-based approach aims to mitigate the potential for errors in tumor demarcation to compound throughout imaging analysis, and thereby improve the consistency of PDAC response assessment to neoadjuvant chemotherapy.

Recently, it has been proposed that image registration may be more robust to noise than the perceptual limits of visualization in the task of identifying low-contrast imaging features.^11,12 Consequently, image registration may more reliably detect or reveal longitudinal changes that would be visually indeterminate or otherwise obscured beneath critical signal-to-noise levels. Registration-based approaches for assessment of chemotherapeutic response have been proposed for cancers of the breast^13–15 and liver^16,17 and have generated promising preliminary data towards their potential utility as novel imaging biomarkers. However, longitudinal registration techniques to analyze treatment response in pancreatic cancer have eluded investigation despite high likelihood for these approaches to be especially impactful due to a paucity of adequate clinical alternatives for treatment monitoring to inform PDAC treatment course.

In this work, two deformable image registration methods are compared for tracking treatment-related changes by measuring the longitudinal dilatation or contraction of a pre-treatment tumor region of interest over the course of neoadjuvant chemotherapy. In 30 patients, changes in lesion size measured by image registration, manual tumor segmentation, RECIST, and pathology-estimated treatment effect are correlated to overall survival (OS) and recurrence-free survival (RFS) to compare registration approaches to conventional clinical variables for tracking longitudinal tumor response. The primary contributions of this work are: (1) we perform the first study on longitudinal image registration of the pancreas with quantitative analysis of target registration errors (TREs), (2) we use deformable image registration to measure changes in the pancreas and tumor volumes to generate new longitudinal imaging biomarkers for PDAC, and (3) we demonstrate the clinical utility of these registration-based biomarkers in predicting OS and RFS in a cohort of 30 patients and compare their predictive power to radiographic and pathologic clinical variables.

2. Methods

For each patient, registration of the pancreas was performed between pre-treatment and post-treatment images obtained at the onset and conclusion of a well-defined neoadjuvant chemotherapy regimen. To this end, two deformable image registration methods were compared, consisting of a biomechanically driven surface-based elastic registration and a symmetrically normalized diffeomorphic intensity-based registration. A block diagram of the registration workflow for measurement of treatment response is shown in Fig. 1. Accuracy of longitudinal registration of the pancreas was evaluated, and a comparison of registration-based, annotation-based, and pathology-based treatment response measures were investigated for association with OS and RFS outcomes.

Fig. 1.

Block diagram of proposed registration-based analysis of treatment response. The pancreas and PDAC tumor are segmented from pre- and post-treatment CT images. The pre- and post- pancreas segmentations are provided to biomechanical contour-based (LIBR) and free-form intensity-based (Greedy SyN) deformable image registration algorithms. The pre-treatment tumor segmentation is morphed according to the transformation field of each registration method to capture changes caused by neoadjuvant therapy. Changes in tumor volume and tumor burden are computed from the deformed tumor region as registration-based measures of longitudinal treatment response. These registration-based measures are explored for association with OS and RFS and compared to measures derived from manual segmentation of the post-treatment tumor, RECIST, and histopathological response.

2.1. Patient Data

Under approval of the institutional review board at Memorial Sloan Kettering Cancer Center, 38 patients were enrolled in a phase II clinical trial (NCT00536874) between July 2007 and December 2011 to evaluate the effect of neoadjuvant gemcitabine and oxaliplatin chemotherapy on outcomes of resectable PDAC.¹⁸ Informed consent was obtained from all patients under the original study and a waiver of Health Insurance Portability and Accountability Act authorization and was granted through the institutional review board at Memorial Sloan Kettering Cancer Center for additional retrospective analysis of these data in accordance with all relevant guidelines and regulations. The patient population consisted of resectable PDAC patients and excluded those with borderline resectable or locally advanced disease. After enrollment, four cycles of gemcitabine and oxaliplatin were administered as neoadjuvant chemotherapy in 2-week intervals. Venous phase contrast-enhanced CT imaging was acquired prior to and following completion of neoadjuvant therapy (henceforth, the pre-treatment and post-treatment images, respectively). After completion of this regimen, those still eligible for resection were operated and continued to receive five additional cycles of adjuvant gemcitabine in 4-week intervals. Image resolution ranged from $0.66\times 0.66\times 2.5{\mathrm{mm}}^3$ to $0.98\times 0.98\times 5.0{\mathrm{mm}}^3$. Radiographic response was assessed by an experienced diagnostic radiologist via RECIST (the percent change in the longest diameters of the target lesions) and from manual segmentations of the pancreas and tumor made in the pre- and post-treatment images. Eight patients were excluded from the original cohort due to missing imaging data. The OS of the remaining 30 patients was tracked over 5 years from the date of first neoadjuvant chemotherapy cycle until death. Twenty-five of 30 patients (83%) continued to resection, and RFS was defined for these 25 resected patients as time from first chemotherapy cycle to the date of clinical diagnosis of local or metastatic recurrence. Five-year follow-up rate was 100% (30/30). Pathological outcomes, including post-resection tumor diameter and estimated treatment effect, were recorded for 22 of 25 resected patients and were missing from electronic records in the remaining three patients.

2.2. Image Registration

First, a biomechanically driven elastic registration was performed using the linearized iterative boundary reconstruction (LIBR) method.¹⁹ This method solves for a set of external forces applied to the organ that generate the observed deformation between pre-treatment and post-treatment pancreas segmentations. Briefly, a finite element model is constructed of the pancreas, and a set of local mechanical perturbations are simulated across the organ in response to a series of control point displacements. Then, a combination of these mechanical perturbations is optimized to reconstruct a volumetric displacement field throughout the organ that minimizes the difference between the post-treatment pancreas shape and a deformed model of the pre-treatment pancreas. It should be noted that this method solely relies on segmentation masks of the outer surface boundary of the pancreas and does not depend on image intensities. Therefore, this method is immune to changes in feature contrast that may occur throughout neoadjuvant therapy due to biological changes in tissue composition. However, information about fine scale motions of tissue features that surround the tumor may be forfeited. The alignment was initialized with a rigid registration optimized by a weighted iterative closest point algorithm.²⁰ The pancreas was simulated as a linear elastic material with stiffness of 2100 Pa, Poisson’s ratio of 0.45, 45 control points, 1.5-mm mesh edge length, and strain energy penalty of $1\times {10}^{-10}{\mathrm{Pa}}^{-2}$.

Second, a diffeomorphic image registration was performed using Greedy.^21,22 This method performs an initial affine registration followed by a symmetrically normalized diffeomorphic optimization to solve for a deformation field that maximizes image similarity on a voxel-by-voxel basis. Greedy is an implementation of the SyN diffeomorphic image registration algorithm in the advanced normalization tools,²³ although with significant runtime optimizations. This image-to-image registration technique is capable of aligning subtle contrast cues around the tumor that may permit more resolved tracking of tumor evolution, although simultaneously the algorithm may become susceptible to variations in intensity contrast profiles that arise from local inflammatory pancreatitis or fibrosis during treatment, abatement or progression of ductal dilation, and variations in intravenous contrast timing between imaging dates. These changes in the pancreas can be substantial and may increase the risk of convergence to local minima during optimization of an image similarity measure. Affine registration was performed using normalized mutual information image similarity, while diffeomorphic registration was performed using normalized cross correlation with $4\times 4\times 4$ voxel radius and multiresolution iteration schedule of $100\times 50\times 10$. Default Gaussian smoothing kernels of ${\sigma }_{\nabla }=\sqrt{3}$ voxels and ${\sigma }_{\mathcal{T}}=\sqrt{0.5}$ voxels were applied to the gradient image and transformation field, respectively, each iteration. Binary image masks of pre- and post-treatment pancreas segmentations were provided to the algorithm to constrain the region of interest and improve overall organ alignment.

After registration, pre-treatment tumor segmentations were deformed according to the computed displacement fields to track the change in tumor from its original shape through the course of neoadjuvant treatment. Representative results from two patients are shown in Fig. 2 for both registration algorithms. Note the subtle and indeterminate contrast boundary of the lesion typical to PDAC in both cases. Relative changes in the pre-treatment tumor volume computed via deformable registration to the post-treatment image are considered to indicate a measure of treatment response. In this paper, validation of pancreas registration accuracy is first performed and then the registration approaches are applied to the prediction of clinical survival outcomes.

Fig. 2.

Pre-treatment segmented pancreas and PDAC lesion (blue) registered to post-treatment pancreas (red) across the course of neoadjuvant chemotherapy for two patients. (a,b) Post-treatment (fixed image) pancreas and tumor of both patients, respectively. (c,d) Rigid registration; (e,f) LIBR registration; and (g,h) Greedy registration of pre-treatment (moving image) to post-treatment pancreas.

2.3. Performance Measures for Registration Accuracy

Accuracy of deformable image registration was evaluated in all 30 patients according to the Dice score coefficient (DSC) of the pancreas, modified Hausdorff distance (MHD), and TRE. The DSC measures the overlap of the registered pre-treatment pancreas with its post-treatment counterpart and is defined as

$$\mathrm{DSC}=\frac{2|\mathcal{T}(A)\cap B|}{|\mathcal{T}(A)|+|B|},$$ (1)

where $\mathcal{T}(A)$ is the registered and transformed pre-treatment pancreas segmentation mask, $B$ is the post-treatment pancreas segmentation mask, and $|·|$ is the set cardinality operator. Distance to agreement between registered pancreas segmentation contours was assessed via the MHD

$$\mathrm{MHD}=\max (\overline{d}(\partial A,\partial B),\overline{d}(\partial B,\partial A)),$$ (2)

where $\partial A$ and $\partial B$ are three-dimensional contours (boundary surfaces) of the pre- and post-treatment pancreas segmentations and $\overline{d}(x,y)$ is the average closest point distance from each point in contour $x$ to contour $y$. Finally, TRE is the gold standard measure for assessing the fidelity of registration approaches by comparing the alignment of corresponding feature points expertly identified in the pre- and post-treatment images. Average TRE was computed via

$$\mathrm{TRE}=\mathrm{mean}({\Vert \mathcal{T}(a)-b\Vert }_2),$$ (3)

where $a$ are manually annotated target positions in the pre-treatment pancreas, $b$ are corresponding manually annotated target positions in the post-treatment pancreas, $\mathcal{T}(·)$ is the transformation field of the registration, and ${\Vert ·\Vert }_2$ is the Euclidean norm. Corresponding points are highly challenging to accurately determine in the pancreas due to scarcity of anatomically consistent high contrast features. Consequently, 4 to 15 (mean 6.9) corresponding targets were selected in each of the pre-treatment and post-treatment images along the pancreatic duct, bile duct, interlobular fat deposits, vascular features, and intrapancreatic cysts when visible. Targets were selected blindly to the registration results and were distributed approximately uniformly throughout the pancreas: 48.6% of all targets were located in either the head or neck of the pancreas and 51.4% of targets were located in either the body or tail of the pancreas. All cases had at least one target in the head or neck of the pancreas and at least one target in the body or tail of the pancreas. Due to inherent limitations of image contrast within the pancreas, it was not always possible to delineate corresponding feature points in the immediate vicinity of the tumor. Figure 3 shows a sample of target points selected on corresponding anatomical features identified in both images.

Fig. 3.

Corresponding intrapancreatic targets identified between pre-treatment and post-treatment imaging. (a) Posterosuperior pancreatoduodenal vein, (b) stent implanted in inferior common bile duct, (c,d) interlobular fatty infiltration, (e) pancreatic cyst centroid, and (f) dilated main duct of pancreas.

2.4. Longitudinal Measures for Response Assessment

In clinical trials, RECIST is the current standard for radiographic assessment of treatment response. This radiographic response is measured from the percent change in the sum of the longest cross-sectional diameters of target lesions between baseline and follow-up imaging. In this study, only a single primary pancreatic tumor was present in each patient. Tumor diameters were manually annotated in the pre- and post-treatment images by an experienced diagnostic radiologist. Pathological variables were also collected for the subset of $n=22$ patients with complete pathology data. These variables included pathology-confirmed post-treatment tumor diameter (PathD) and pathological treatment effect (Path%), defined as the estimated percentage of tumor cell necrosis on histologic evaluation of the resected specimen.^24,25

In addition to these clinical comparators, this work introduces a registration-based imaging biomarker approach to quantify longitudinal treatment response of the tumor via two proposed metrics. The first metric measures the percent change in tumor volume from pre-treatment to post-treatment previously employed by Perri et al.,¹⁰ defined as

$$V=\frac{v_{\mathrm{post}}^T-v_{\mathrm{pre}}^T}{v_{\mathrm{pre}}^T}\times 100\%,$$ (4)

where $v_{\mathrm{pre}}^T$ is the volume of the pre-treatment region identified as tumor and $v_{\mathrm{post}}^T$ is the deformed volume of the identified tumor region after registration of pre-treatment to post-treatment images. The second metric measures the change in tumor burden within the pancreas

$$B=(\frac{v_{\mathrm{post}}^T}{v_{\mathrm{post}}^P}-\frac{v_{\mathrm{pre}}^T}{v_{\mathrm{pre}}^P})\times 100\%,$$ (5)

where $v_{\mathrm{pre}}^P$ is the volume of the pre-treatment pancreas and $v_{\mathrm{post}}^P$ is the volume of the post-treatment pancreas. These measures were computed based on the deformed tumor volumes from LIBR and Greedy image registration ($V_{\mathrm{L}\mathrm{I}\mathrm{B}\mathrm{R}}$ and $B_{\mathrm{L}\mathrm{I}\mathrm{B}\mathrm{R}}$, $V_{\text{Greedy}}$, and $B_{\text{Greedy}}$, respectively). The post-treatment tumor volumes were also segmented from the post-treatment image to assess the registration-based response metrics when relying on expertly segmented manual measurements instead of an image registration workflow ($V_{\text{Manual}}$ and $B_{\text{Manual}}$).

2.5. Statistical Analysis

Each radiographic and pathologic tumor response measure was evaluated for significance as a predictor variable for OS and RFS with a continuous univariate Cox proportional hazards model, with hazard ratio (HR) as the primary endpoint. Correlations of response measures to OS and RFS event times were characterized by Harrell’s C-index. To inspect temporal performance, binary prediction of 1-, 2-, 3-, and 5-year OS and RFS was also quantified by the area under the receiver operator characteristic curve (AUC), with univariate statistical analysis carried out with Wilcoxon rank sum tests between surviving and non-surviving groups. Prognostic utility of the proposed registration approach was assessed by comparing separation of survival curves between high and low risk groups partitioned by median values of RECIST score, registration-based, and manual segmentation-based measures of treatment response via Kaplan-Meier statistics and the log rank test. For analysis of registration errors, DSC, MHD, and TRE measures were compared via Wilcoxon rank sum tests. Statistical significance was considered at a level of $\alpha =0.05$.

3. Results

3.1. Registration Error

Registration errors for each case are displayed in Fig. 4. For LIBR and Greedy registration, respectively, mean and standard deviation (median) of DSC was $0.90\pm 0.05$ (0.92) and $0.92\pm 0.07$ (0.93), MHD was $0.61\pm 1.26$ (0.24) mm and $0.60\pm 1.67$ (0.23) mm, and TRE was $2.56\pm 2.45$ (2.05) mm and $4.15\pm 3.63$ (3.30) mm across all cases. In comparison, values of DSC, MHD, and TRE of iterative closest point rigid registration were $0.78\pm 0.08$ (0.80), $2.75\pm 1.31$ (2.46) mm, and $4.77\pm 2.40$ (4.15) mm, respectively. Both deformable registration methods were found to significantly improve over rigid registration according to each metric ($p<0.05$). LIBR registration was associated with significantly lower average TRE than Greedy registration ($p<0.001$), while Greedy registration was associated with significantly higher Dice overlap than LIBR registration ($p<0.001$). No significant difference was found in median MHD between LIBR and Greedy registration methods ($p=0.57$). Considering the axial spacing of images, the average magnitude of TRE was on the order of 1–2 voxels for both deformable image registration methods. One case seen in Fig. 4 was associated with higher registration errors due to local misregistration arising due to inconsistent segmentation of the organ boundary between the head of the pancreas and the duodenum. This misregistration was distant to the tumor and did not affect local tumor alignments. Registration errors associated with this case were included in the overall measures of registration accuracy to reflect performance under a realistic, clinically viable registration workflow using a uniform set of registration parameters without additional performance tuning for each case.

Fig. 4.

Pancreas registration accuracy (DSC, MHD, and TRE) in 30 patients for iterative closest point method (rigid), biomechanical elastic method (LIBR), and diffeomorphic SyN method (Greedy). * ($p<0.001$).

3.2. Survival Analysis

Median OS of the cohort was 31 months (95% CI: 19 to 42) while median RFS was 17 months (95% CI: 12 to 28) (Fig. 5). Table 1 (OS) and Table 2 (RFS) report the hazard ratio and $p$-value of univariate Cox regression and C-index of tumor response metrics to survival outcomes. The only predictors of OS (Table 1) with HRs substantially greater than one were the registration-predicted differences in tumor burden $B_{\text{Greedy}}$ (HR = 1.53) and $B_{\text{LIBR}}$ ($\mathrm{HR}=1.91$). While $B_{\text{Greedy}}$, $B_{\text{LIBR}}$, $V_{\text{Greedy}}$, and RECIST were all associated with C-index greater than 0.60, the HRs for OS of $V_{\text{Greedy}}$ and RECIST were only marginally greater than unity (HR = 1.01). ${\mathit{B}}_{\text{Greedy}}$ and RECIST were significantly associated with OS ($p<0.01$ and $p<0.05$, respectively), while $B_{\text{LIBR}}$ and $V_{\text{Greedy}}$ were nearly significant ($p=0.079$ and $p=0.059$, respectively). Although the change in tumor burden measured with Greedy image registration was a significant predictor of OS, this quantity measured with LIBR registration was not found to reach statistical significance but was associated with the highest HR. Importantly, neither the changes in tumor volume $V_{\text{Manual}}$ or tumor burden $B_{\text{Manual}}$ measured from manual segmentation of pre-and post-treatment tumor volumes, nor the pathological variables PathD or Path% were found to be strong predictors of OS. Similar trends are reflected in prediction of RFS (Table 2), where only $B_{\text{Greedy}}$, $B_{\text{LIBR}}$, RECIST, and PathD were associated with C-index $>0.60$, and only ${\mathit{B}}_{\text{Greedy}}$ and ${\mathit{B}}_{\text{LIBR}}$ were statistically significant predictors of RFS ($p<0.05$) with HRs of 1.39 and 2.53, respectively. No other radiographic or pathologic response measures, nor the tumor burden measure $B_{\text{Manual}}$ assessed by manual segmentation of pre- and post-treatment images, were significantly associated with RFS. These findings suggest that registration-based measures of change in tumor burden, i.e., the difference in the relative proportion of tumor volume to total organ volume, may be a more clinically relevant indicator of disease response in PDAC than tumor-centric size measures. The limited performance of the manual segmentation approaches towards predicting outcome is also consistent with challenges in visual delineation of PDAC leading to measurement variabilities that likely obscure underlying treatment effects.

Fig. 5.

Kaplan-Meier survival curves for OS (black) and RFS (red). Number of patients at risk displayed above horizontal axis.

Table 1.

Predictors of OS.

Metric	Cox HR (95% CI)	Cox $p$	C-index	AUC-1	AUC-2	AUC-3	AUC-5
$V_{\text{LIBR}}$	1.01 (0.98–1.03)	0.678	0.55	0.85*	0.60	0.52	0.43
$V_{\text{Greedy}}$	1.01 (1.00–1.02)	0.059	0.60	0.88*	0.69	0.57	0.54
$V_{\text{Manual}}$	1.00 (0.99–1.01)	0.856	0.48	0.75	0.57	0.37	0.23
$B_{\text{LIBR}}$	1.91 (0.93–3.93)	0.079	0.63	0.85*	0.55	0.68	0.65
$B_{\text{Greedy}}$	1.53 (1.15–2.03)	0.004	0.64	0.84*	0.72*	0.66	0.69
$B_{\text{Manual}}$	0.95 (0.80–1.13)	0.567	0.42	0.60	0.43	0.38	0.32
RECIST	1.01 (1.00–1.03)	0.036	0.62	0.91*	0.73*	0.60	0.41
PathD	1.13 (0.75–1.69)	0.558	0.58	0.89	0.68	0.65	0.46
Path%	0.99 (0.97–1.01)	0.296	0.47	0.41	0.42	0.53	0.56

$V$: percent change in tumor volume; $B$: difference in tumor burden; PathD: pathology-confirmed post-treatment diameter; Path%: pathology-estimated treatment effect; and HR: hazard ratio. Significant predictors emphasized in bold.

^*Asterisk denotes $p<0.05$ of Wilcoxon rank-sum test.

Table 2.

Predictors of RFS.

Metric	Cox HR (95% CI)	Cox $p$	C-index	AUC-1	AUC-2	AUC-3	AUC-5
$V_{\text{LIBR}}$	1.00 (0.97–1.02)	0.713	0.50	0.61	0.45	0.44	0.33
$V_{\text{Greedy}}$	1.01 (0.99–1.02)	0.338	0.53	0.58	0.54	0.56	0.48
$V_{\text{Manual}}$	1.00 (0.99–1.01)	0.923	0.46	0.59	0.42	0.33	0.23
$B_{\text{LIBR}}$	2.53 (1.01–6.32)	0.048	0.63	0.64	0.60	0.78*	0.75
$B_{\text{Greedy}}$	1.39 (1.01–1.92)	0.044	0.60	0.60	0.60	0.75	0.73
$B_{\text{Manual}}$	0.93 (0.77–1.12)	0.418	0.43	0.47	0.38	0.33	0.29
RECIST	1.02 (0.99–1.04)	0.289	0.61	0.90*	0.59	0.45	0.33
PathD	1.06 (0.72–1.54)	0.779	0.64	0.70	0.65	0.54	0.42
Path%	0.99 (0.97–1.01)	0.318	0.49	0.41	0.54	0.53	0.55

$V$: percent change in tumor volume; $B$: difference in tumor burden; PathD: pathology-confirmed post-treatment diameter; Path%: pathology-estimated treatment effect; and HR: hazard ratio. Significant predictors emphasized in bold.

^*Asterisk denotes $p<0.05$ of Wilcoxon rank-sum test.

To evaluate temporal performance, AUC for binary prediction of 1-, 2-, 3-, and 5-year OS and RFS are also shown in Tables 1 and 2, respectively, with significant differences between median predictor values of surviving and non-surviving groups marked (*) where $p<0.05$ by Wilcoxon rank-sum test. Although RECIST score and change in tumor volume predicted by image registration were found to be strong predictors of early outcome (1-year and 2-year OS), these methods performed less effectively over longer horizons of 3-year and 5-year survival. Registration-derived changes in tumor burden offered the most consistent predictive performance across all intervals of OS and RFS. Conversely, changes in tumor volume and tumor burden determined by comparisons of manual tumor segmentations were found to be the worst performing measurements of treatment response in relation to survival prediction. Although there is room for improvement, no effective non-invasive longitudinal assessment strategies for PDAC yet exist. The AUC values reported here compare favorably to current clinical staging systems for PDAC, which have been shown to predict 1-, 2-, and 3-year OS with AUC of merely 0.6.²⁶

Pathology-confirmed tumor size (PathD) and estimated treatment effect (Path%) determined from post-resection tissue samples were not found to correlate significantly with OS (Table 1) or RFS (Table 2). However, post-treatment tumor diameters measured from ground-truth pathology were associated with a comparable RFS C-index. The observed trends suggest that the non-invasive registration-based tumor burden metrics convey stronger predictive signal towards survival outcomes than the individual radiological or pathological clinical variables.

To identify prognostic value of the proposed registration-derived measures for tumor response, patients were divided into high (+) and low (–) risk groups based on the median values of each radiographic response assessment technique. The median value was selected as a cutoff to produce groups of equal size to maximize statistical power in two-class comparison of survival outcomes across worst and best responders stratified by the proposed response measures. In Fig. 6, no significant difference was found between OS or RFS in patients above (RECIST+) and below (RECIST−) the median RECIST scores in the cohort. Patients separated by the median change in tumor burden measured by LIBR registration ($B_{\text{LIBR}}+$ versus $B_{\text{LIBR}}-$) showed qualitative but nonsignificant separation of OS ($p=0.06$) and RFS ($p=0.10$) curves. However, separation according to the median change in tumor burden by Greedy registration ($B_{\text{Greedy}}+$ versus $B_{\text{Greedy}}-$) was associated with significant separation in OS ($p<0.05$) and RFS ($p<0.01$). These results suggest that treatment response assessed by longitudinal image registration could index survival outcomes more effectively than the percent change in diameters defined in RECIST guidelines. Of note, $B_{\text{Manual}}$ stratified survival outcomes in the opposite direction than expected, where patients in the low risk group were counterintuitively associated with non-significantly shorter OS ($p=0.12$) and RFS ($p=0.12$) outcomes. This behavior is consistent with Tables 1 and 2, where ${\mathit{B}}_{\text{Manual}}$ tended to incorrectly predict OS and RFS with 2-year, 3-year, and 5-year AUC values below 0.5, signifying that manual segmentation may yield predictors that are inferior to random chance. Therefore, registration-based approaches could stratify patient outcomes in PDAC more effectively than manual annotation-based imaging assessments commonly regarded in clinical practice.

Fig. 6.

Kaplan-Meier survival curves (top: OS; bottom: RFS) comparing outcomes from the set of patients in the low risk (solid line) versus high risk (dashed line) groups of (a) RECIST score, (b) change in tumor burden by LIBR registration, (c) change in tumor burden by Greedy registration, and (d) change in tumor burden by manual segmentations. Statistical results via log rank test.

4. Discussion

Current treatment response measures that rely on manual annotation may become less congruent when considering potential for perceptual inconsistencies to arise from the challenging imaging characteristics of PDAC. Precise tumor delineation is largely impeded by poor lesion contrast and diffuse lesion boundaries typical to the radiographic presentation of this tumor type. This study shows that treatment response measured by longitudinal image registration can predict survival outcomes more effectively than measures based on manually repeated diametric annotations, volumetric segmentations, or pathology readings of PDAC lesions. Of note, agreements between PathD and radiographic diameters of the tumor were associated with coefficients of accuracy²⁷ of 0.99 and 0.98 for LIBR and Greedy registrations of the pre-treatment tumor segmentation to the post-treatment image, which signifies low overall bias in either underestimating or overestimating the tumor diameter when factoring out measurement variance. However, the manually segmented post-treatment tumor diameters were associated with lower accuracy of 0.86, suggesting that manual tumor delineation in post-treatment images may become biased due to treatment-related changes that further obfuscate already ill-defined tumor borders. Treatment-related effects including encapsulating fibrosis and pancreatitis could make accurate assessment of tumor extent more challenging in post-treatment compared to pre-treatment images.^28,29 On the contrary, the registration-based approach for assessing longitudinal change may mitigate unintended measurement biases by circumventing the need to make equally accurate tumor measurements in both pre-treatment and post-treatment images.

While pathology measures are commonly accepted as a clinical gold standard, it is important to stress that these measures are obtained after resection and cannot be used prognostically. Moreover, the PathD tumor size at the time of resection does not comprehensively encompass disease state. Interval changes across the chemotherapeutic window are likely to provide more targeted information about patient-specific differences in susceptibility to treatment, underlying disease aggressiveness, and parenchymal adaptations that correlate to survival outcomes. Yet, pathological measures of treatment effect (Path%) tend to diminish in predictive strength in patients with incomplete therapeutic response. Although it has been established that Path% is effective at stratifying survival outcomes between patients who achieve pathological major response ($<10\%$ viable residual tumor) and those who do not,³⁰ there is less evidence to support its utility in stratifying survival outcomes among patients who achieve less than major response.³¹ In the cohort investigated herein, only one patient achieved pathological major response and no patients achieved pathological complete response. The present study agrees with the findings of Ref. 28, which similarly identified that neither pathological tumor size nor treatment effect were significantly associated with survival among low-to-moderate grade responders.

Of the radiographic biomarkers, the registration-predicted changes in tumor volume were only weakly correlated with RECIST score, with Spearman’s $\rho$ of 0.32 and 0.24 for $V_{\text{LIBR}}$ and $V_{\text{Greedy}}$, respectively. Furthermore, changes in tumor burden exhibited even weaker correlations with RECIST ($\rho =0.23$ and 0.09 for $B_{\text{LIBR}}$ and $B_{\text{Greedy}}$, respectively). These findings suggest that the registration-based tumor measures may serve as semi-independent predictors to the clinical radiographic standard. Moreover, the tumor burden measure, which incorporates volumetric parenchymal changes, likely provides additional information that is not captured by tumor-centric measures, such as effects of parenchymal atrophy that are often associated with disease progression. Beyond these differences, tumor response measures based on manual segmentations of the pre- and post-treatment tumor were not found to be predictive of survival outcomes, indicating the importance of the registration process towards harmonizing measurement variabilities across the treatment interval and achieving more consistent longitudinal response measures. Response evaluations measured through tumor burden as opposed to tumor volume and longitudinal image registration as opposed to manual tumor segmentation therefore may lead to novel imaging biomarkers for therapeutic response in PDAC. This registration-based approach replaces the need for manual lesion annotations that are susceptible to compounding measurement errors with a computational analysis workflow that depends only on outer pancreas segmentation masks and an initial region of interest around the tumor that could reasonably be obtained by automatic methods.

Although neither registration algorithm was specifically designed or optimized for registration of the pancreas, both methods produced significant predictors of clinical outcomes. While LIBR registration achieved significantly lower TRE throughout the pancreas than Greedy registration, Greedy registration achieved significantly higher DSC. It is important to note that while the LIBR registration elastically parameterizes organ deformation to a reduced optimization space defined over a sparse set of control points on the organ boundary, the Greedy algorithm incorporates substantially more degrees of freedom within a free-form deformation basis to approximate alignments from local intensity information. The difference in TRE is likely attributable to the biomechanical basis of the LIBR registration providing physical consistency to the deformation field, while the hyper-local control afforded to deformations in the Greedy algorithm may provide greater flexibility in fine-tuning DSC overlap but risk convergence to local minima in the presence of longitudinal shifts in pancreatic texture, treatment effects, contrast timing, and volume averaging artifacts. Although the additional degrees of freedom in the Greedy algorithm may achieve more detailed local estimates of nonelastic peri-tumor effects when internal contrast and texture information within the pancreas is sufficient for convergence to accurate alignments, the LIBR algorithm nonetheless was associated with higher HRs for prediction of OS and RFS outcomes in Tables 1 and 2. This behavior could be attributable to the underlying biomechanical model specifically filtering for evidence of bulk interval mass effect from subtle changes in pancreas caliber. The elevated HRs associated with $B_{\text{LIBR}}$ over $B_{\text{Greedy}}$ may even suggest that the biomechanically elastic constraint could have a stabilizing effect on the ability to detect treatment response or progression. While there are merits to the distinct capabilities of each registration approach, the two methods occupy opposing sides of the bias-variance tradeoff, which could be rebalanced by a mixed approach that incorporates biomechanical information or regularization into the intensity-based registration, by developing registration techniques to more explicitly model physiological changes in the pancreas, or further by incorporating tumor growth models to specifically parameterize tumor response. These potential algorithmic advancements will be important directions of future research. Yet, the approach that we propose adopts non-customized image registration algorithms that offer good runtime performance and ease of implementation readily achievable within the clinical workflow.

Despite the promising predictive performance of the proposed registration-based biomarkers, survival outcomes also likely depend on underlying disease biology, which may not be fully encompassed by morphometric imaging responses to neoadjuvant therapy. Comprehensive prediction of PDAC survival outcomes remains a highly challenging task due to variations in tumor phenotype and aggressiveness that are rarely known at the time of treatment. A multi-effect model that integrates these findings with other emerging and established non-invasive imaging, biochemical, and genomic biomarkers, such as radiomic features, CA 19-9, and high-throughput sequencing may further improve overall prediction and lead to new clinical models for disease restaging. However, larger cohorts are needed to establish fine-tuned cutoffs for the proposed registration-based biomarkers to achieve optimal patient stratification.

This analysis is limited by evaluation on a small sample of narrowly selected patients who were deemed initially eligible for resection and enrolled in a specific treatment regimen that included neoadjuvant chemotherapy. Further verification is needed in a larger cohort of more broadly selected patients with greater sample size to establish the levels of certainty, reproducibility, and reliability required to establish clinically acceptable imaging biomarkers. Yet, association of the proposed registration measures to survival outcomes is nonetheless successfully demonstrated in the present study with sufficient power to detect significant HRs associated with the effect sizes of the registration-based measures. Considering the lack of effective clinical prognostic models for PDAC, current radiographic and pathologic measures available in clinical practice are likely associated with smaller effect sizes that fail to significantly associate with survival outcomes in all but the most strongly powered studies.

5. Conclusion

This study presents a longitudinal image registration approach for evaluating response to neoadjuvant therapy in resectable PDAC that may predict OS and RFS more effectively than current clinical univariate predictors over 5-year follow-up. Pancreas registration was shown to be accurate within errors of 1–2 voxels on average. These methods provide a unique approach for response assessment that aims to overcome limitations associated with measurement of therapeutic response in PDAC tumors, which are frequently associated with poor radiographic contrast. These advances could contribute imaging biomarkers to improve prediction and stratification of treatment responses to better inform clinical management of pancreatic cancer.

Biographies

Jon S. Heiselman, PhD, received his BSE degree in biomedical engineering from the University of Michigan in 2015 and his PhD in biomedical engineering from Vanderbilt University in 2020. He is currently a postdoctoral researcher at Memorial Sloan Kettering Cancer Center with joint appointment at Vanderbilt University. His research interests include image registration, biomechanics, and computational modeling to improve preoperative assessment, intraoperative delivery, and postoperative monitoring of therapy. He has been a member of SPIE since 2017.

Brett L. Ecker, MD, received his medical degree with a distinction in research at Mount Sinai School of Medicine; General Surgery training at the University of Pennsylvania; and Complex General Surgical Oncology fellowship at Memorial Sloan Kettering Cancer Center, where he acted as the chief fellow for the 2021–2022 academic year. Currently, he is an assistant professor of surgery at Rutgers Cancer Institute of New Jersey.

Michael I. Miga, PhD, received his PhD from Dartmouth College, specializing in biomedical engineering.  He joined the faculty in the Department of Biomedical Engineering at Vanderbilt University in 2001 and is the Harvie Branscomb Professor at Vanderbilt and a professor of Biomedical Engineering. He is director of the Biomedical Modeling Laboratory, and co-founder of the Vanderbilt Institute for Surgery and Engineering.  He is also PI and director of a novel NIH T32 training program entitled “Training Program for Innovative Engineering Research in Surgery and Intervention” that is focused on the creation of translational technologies for treatment and discovery in surgery and intervention. He also was a co-inventor of the first FDA cleared image guided liver surgery system. He is an AIMBE and SPIE fellow.  His research interests are in computational modeling, inverse problems/computational imaging, soft-tissue biomechanics/biotransport, technology-guided therapy, image/imaging-guided surgery and intervention, and data-driven procedural medicine.

William R. Jarnagin, MD, has been an attending surgeon at Memorial Sloan-Kettering Cancer Center since 1997, where he has served as a chief of the Hepatopancreatobiliary Service since 2008 and was a vice-chairman of the Department of Surgery from 2006 to 2010. He holds the Leslie H. Blumgart, M.D. Chair in Surgical Oncology and is professor of Surgery at Weill Medical College of Cornell University.

Natally Horvat, MD/PhD, received her MD degree from Faculdade de Medicina de Campos, Rio de Janeiro, Brazil, in 2009; radiology residency and fellowship from the University of São Paulo in 2014; clinical research fellowship in Oncologic Imaging from Memorial Sloan Kettering Cancer Center in 2017; and PhD in radiology from the University of São Paulo in 2019. Currently, she is working as an assistant attending in the Department of Radiology at Memorial Sloan Kettering Cancer Center, assistant professor of Radiology at Weill Medical College of Cornell University, and visiting professor at the University of São Paulo. Her research is focused on improving the accuracy of imaging modalities in the diagnosis and risk stratification of abdominal malignancies. She is at the forefront of discovering qualitative and quantitative imaging biomarkers. She collaborates in multidisciplinary groups to construct prediction models that anticipate relevant patient outcomes. By integrating radiological data into multimodality approaches, she aims to develop sustainable personalized medicine strategies for diagnosis, patient stratification, and treatment assessment.

Jayasree Chakraborty, PhD, is an assistant attending in the Department of Surgery at Memorial Sloan Kettering Cancer Center. She received her PhD in computer vision and machine learning from IIT Kharagpur in 2013 and served as an assistant professor at NIT Silchar from 2013 to 2015. Her research area is focused on the development of AI-based models to detect, diagnose and outcome prediction for different cancers. She has more than 90 publications in journals and conferences of international repute.

Biographies of the other authors are not available.

Disclosures

Dr. O’Reilly reports research funding from Genentech/Roche, Celgene/BMS, BioNTech, AstraZeneca, Arcus, Elicio, and Parker Institute, and additional conflicts of interest involved with data safety and monitoring board (DSMB) and consulting activity with Cytomx Therapeutics (DSMB), Rafael Therapeutics (DSMB), Seagen, Boehringer Ingelheim, BioNTech, Ipsen, Merck, IDEAYA, Silenseed, Novartis, AstraZeneca, BioSapien, Cend Therapeutics, Astellas, Thetis, Autem, Novogene, ZielBio, Tempus, Fibrogen, Agios, Genentech-Roche, Eisai, and Zymeworks. The other authors have no relevant financial or non-financial conflicts of interests to disclose.