Prediction of neoplastic progression in Barrett’s esophagus using nanoscale nuclear architecture mapping: a pilot study

Abstract

Background and Aims:

Nanoscale nuclear architecture mapping (nanoNAM), an optical coherence tomography–derived approach, is capable of detecting with nanoscale sensitivity structural alterations in the chromatin of epithelial cell nuclei at risk for malignant transformation. Because these alterations predate the development of dysplasia, we aimed to use nanoNAM to identify patients with Barrett’s esophagus (BE) who might progress to high-grade dysplasia (HGD) or esophageal adenocarcinoma (EAC).

Methods:

This is a nested case-control study of 46 BE patients, of which 21 progressed to HGD/EAC over 3.7 ± 2.37 years (cases/progressors) and 25 patients who did not progress over 6.3 ± 3.1 years (control subjects/non-progressors). The archived formalin-fixed paraffin-embedded tissue blocks collected as part of standard clinical care at the index endoscopy were used. nanoNAM imaging was performed on a 5-μm formalin-fixed paraffin-embedded section, and each nucleus was mapped to a 3-dimensional (3D) depth-resolved optical path difference (drOPD) nuclear representation, quantifying nanoscale-sensitive alterations in the 3D nuclear architecture of the cell. Using 3D-drOPD representation of each nucleus, we computed 12 patient-level nanoNAM features summarizing the alterations in intrinsic nuclear architecture. A risk prediction model was built incorporating nanoNAM features and clinical features.

Results:

A statistically significant differential shift was observed in the drOPD cumulative distributions between progressors and nonprogressors. Of the 12 nanoNAM features, 6 (mean-maximum, mean-mean, mean-median, entropy-median, entropy-entropy, entropy-skewness) showed a statistically significant difference between cases and control subjects. NanoNAM features based prediction model identified progression in independent validation sets, with an area under the receiver operating characteristic curve of 80.8% ± .35% (mean ± standard error), with an increase to 82.54% ± .46% when combined with length of the BE segment.

Conclusions:

NanoNAM can serve as an adjunct to histopathologic evaluation of BE patients and aid in risk stratification.

The prevalence of Barrett’s esophagus (BE), the precursor for esophageal adenocarcinoma (EAC), is estimated to be .96% in general population and 6.72% in patients with GERD.¹ Although certain factors such as male gender, obesity, smoking, length of BE segment, and hiatal hernia influence the risk of neoplastic progression in BE, the current recommendations in the management of BE rest solely on the presence of dysplasia.² Most society guidelines recommend surveillance endoscopies every 3 to 5 years in nondysplastic BE (NDBE) and endoscopic eradication therapy in confirmed cases of low-grade dysplasia (LGD) and high-grade dysplasia (HGD).³ However, dysplasia is not an entirely objective marker for neoplastic progression in BE. First, dysplasia as a histologic marker is highly susceptible to interobserver variability even among expert GI pathologists.⁴ In a study of 7 expert pathologists evaluating 72 slides of BE patients with varying degrees of dysplasia, the κ value was .22 (95% confidence interval [CI], .11-.29) for NDBE, .11 (95% CI, .004-.15) for LGD, and .43 (95% CI, .36-.46) for HGD.⁴ Another factor is the unpredictable progression of dysplasia to cancer. EAC is hypothesized to occur through a stepwise progression from histopathologic stages of intestinal metaplasia to LGD to HGD to intramucosal EAC and finally to invasive EAC.⁵ However, this is not observed in real-world practice. Studies show that the most common histologic diagnosis on an endoscopy before development of EAC is NDBE.^6,7 On the other hand, most NDBE patients do not develop cancer. The risk of progression to EAC in NDBE is only .12% to .14% per year.^8,9 Even in LGD, up to 60% have no dysplasia on follow-up endoscopies.¹⁰ Therefore, the ideal scenario in the management of NDBE is identifying patients who are at high risk for progression to cancer and offering endoscopic eradication therapy to them while reducing the frequency of surveillance in patients at low risk of progression. Risk prediction has many advantages: reducing unnecessary endoscopies in low-risk patients and early intervention in high-risk patients, ultimately leading to reduced costs of management of BE. Therefore, identification of an ideal marker for risk stratification of neoplastic progression in BE is a subject of active research.

Several risk prediction models based on clinical factors, endoscopic findings, and biomarkers have been proposed.^11-14 For example, several patient factors such as older age, smoking, BE segment length, nodularity, and genetic markers such as p53 mutation, TP53 loss of heterozygosity, CDKN2A loss of heterozygosity, tetraploidy, abnormal DNA ploidy, and expression of Aspergillus oryzae lectin have been reported to be associated anywhere from a 3- to 38-fold increased risk of neoplastic progression in BE.² However, some of them have not been externally validated and are not currently recommended for identifying patients who might require early intervention from those who can be managed by extended surveillance intervals.

In the above context and given that histologic diagnosis continues to be the criterion standard in BE management, approaches that can aid in better identification of NDBE patients at high risk of progression to EAC will help improve management of BE. We developed an optical imaging approach that can objectively quantify abnormal architectural alterations in the nuclei of nondysplastic epithelial cells in NDBE patients with nanoscale sensitivity.¹⁵ Our approach, which we refer to as nanoscale nuclear architecture mapping (nanoNAM), has previously shown the potential to capture subtle neoplastic transformations that are hard to visualize at the microscale during histologic diagnosis of precancerous samples.¹⁵ We hypothesize that nanoNAM can identify nuclear architectural changes before the development of neoplasia in NDBE patients and therefore help to differentiate progressors from nonprogressors. Therefore, we propose a proof-of-concept study using nanoNAM-based optical markers in NDBE epithelium to identify patients at risk of developing neoplasia. We also aimed to build a prediction model with nanoNAM-based optical markers as input to a priori classify NDBE patients into progressors and nonprogressors.

METHODS

Patient selection and histopathologic characterization

The Cleveland Clinic BE Registry is a prospectively collected database of patients with BE seen at the Cleveland Clinic and affiliated hospitals. Study patients were identified from our BE registry. This is a nested case-control study of NDBE patients who progressed to HGD or EAC at least 1 year or more from the index endoscopy (cases/progressors) and NDBE patients who did not develop any dysplasia or cancer on ≥3-year follow-up (control subjects/nonprogressors).

Inclusion criteria were availability of adequate tissue and clinicopathologic data and subspecialist review of histologic grade of dysplasia. Clinical variables collected were age, gender, race, smoking, alcohol use, body mass index, BE length, hiatal hernia size, and presence and type of visible lesions. Cases and control subjects were matched for age, gender, smoking history, and age of the tissue sample. Patients enrolled in the surveillance program underwent endoscopies at recommended intervals per guidelines and had 4-quadrant biopsy sampling done every 1- to 2-cm intervals based on the Seattle protocol.¹⁶

Among cases, follow-up was defined as the time difference between index EGD and EGD revealing HGD/cancer. In the control group, follow-up was defined as the time between the index endoscopy and the latest endoscopy. Control patients did have interim surveillance endoscopies as per prevalent guidelines at that time.

Histologic grading was assigned per the Vienna classification into no dysplasia, indefinite for dysplasia, LGD, HGD, intramucosal carcinoma, and invasive cancer.¹⁷ For study purposes, formalin-fixed paraffin-embedded tissue blocks from the baseline endoscopies were retrieved and sectioned at 5-μm thickness using a microtome. The unstained sections were placed on coated glass slides and imaged using nanoNAM instruments to obtain the depth-resolved optical path difference (drOPD) measurements.¹⁵ The slides were reviewed by an expert GI pathologist to confirm the cell nuclei imaged using nanoNAM was histologically normal and to grade the inflammation.

Two patients in the progressor group and 1 patient in the nonprogressor group had erosive esophagitis at the index EGD. In all 3 patients, the biopsy sampling site was distant from the area of inflammation. All patients were on proton pump inhibitor therapy at the time of EGD. All cases of dysplasia were confirmed by a second GI pathologist or at a pathology consensus conference.

Principles of nanoNAM

nanoNAM is a Fourier-domain optical coherence tomography–derived approach that uses common-path interferometry to capture the phase from the backscattered signal from unstained BE tissue samples.¹⁸ This is done using a trimodal system that uses a 2-stage workflow to reliably identify epithelial cell nuclei on label-free BE tissue samples. These signals are analyzed to compute 3-dimensional (3D), nanoscale-sensitive, intrinsic nuclear architecture properties of these epithelial cells associated with depth-resolved alteration in mean nuclear optical density and depth-resolved localized heterogeneity in optical density of the cell nuclei. Detailed description of the optical system along with the computational analysis can be found elsewhere.¹⁵ Here, “depth-resolved” emphasizes the ability of nanoNAM to capture intrinsic nuclear architecture properties at each coherence-gated optical depth location along the entire optical thickness of the tissue section.

nanoNAM-based sample characterization

Using nanoNAM, we mapped each nucleus to a drOPD value quantifying 3D nuclear architecture of the cell along the lateral (x, y) and axial (z) directions.^15,18 Using the drOPD value of each nucleus, we computed 12 patient-level nanoNAM features that summarized the statistical distribution and heterogeneity of the nanoNAM-based quantification of intrinsic nuclear structure: mean, median, maximum, standard deviation, skewness (asymmetry), and entropy (heterogeneity) of drOPD values of all nuclei across the lateral direction at axial depths that were 45 nm apart. Next, we calculated mean and entropy in the axial direction resulting in the following 12 patient-level drOPD statistics: mean-mean, mean-median, mean-maximum, mean-standard deviation, mean-skewness, mean-entropy, entropy-mean, entropy-median, entropy-maximum, entropy-standard deviation, entropy-skewness, and entropy-entropy. These 12 features were compared between the cases and control subjects.

Support vector machine–based prediction of BE patients progressing to neoplasia

Based on the differences in nanoNAM features and clinical features between progressors and nonprogressors, we built a prediction model using a support vector machine (SVM)-based classifier.¹⁹ We trained a nu-parameterized soft-margin SVM to classify NDBE patients into cases or control subjects based on the nanoNAM features and clinical characteristics that demonstrated a statistically significant difference between the 2 groups. Soft-margin–based learning optimally manages the trade-off between the classification ability of SVM and noise in the data. This trade-off is quantified by the nu-parameter, which is bounded between 0 and 1.²⁰ We split the NDBE patient cohort into mutually exclusive training and validation sets using stratified sampling to ensure that both sets, respectively, contained 50% of the relative proportion of progressor and nonprogressor BE patients, ensuring a similar relative mix of 2 groups in the 2 mutually exclusive sets. We trained the SVM using leave-1-out cross-validation and tested it on the independent validation set. We used the area under the receiver operating characteristic (AUROC) curve as the metric for quantifying testing performance. To test robustness of performance, we combined stratified sampling with bootstrapping to generate 500 pairs of training and validation sets. SVM was trained using each training set and tested on its corresponding validation set. The resulting AUROC metric was used to compare the performance between different input features ranging from nanoNAM features, clinical features, to a combination of both.

RESULTS

Study population

There were 21 progressors (16 to HGD and 5 to EAC) during a mean follow-up of 3.7 years. The baseline characteristics and demographics are presented in Table 1. Except for longer BE segment in progressors, the other features were comparable between both groups. Of 21 progressors, 12 had at least a 3-year difference between the 2 endoscopies. In the remaining progressors, reasons for follow-up EGD <3 years were dysphagia (n = 2), GI bleed (n = 1), persistent reflux (n = 1), intervening indefinite for dysplasia or LGD (n = 3), and ordered by their primary care provider (n = 2). Before 2011, the recommendation per the guidelines was 2 EGDs with biopsy sampling within 1 year for documentation followed by EGDs every 3 years.

TABLE 1.

Baseline characteristics and demographics of cases and control subjects

Factor	Total subjects (n = 46)	Cases (progressors) (n = 21)	Control subjects (nonprogressors) (n = 25)	P value
Gender
Male	34 (73.91)	17 (80.96)	17 (68)	.5*
Female	12 (26.09)	4 (19.04)	8 (32)
Race
White	45	20	25	.46*
Black	1	1	0
Other	0	0	0
Age of the subject (date of initial endoscopy minus date of birth), y		60.0 ± 14.55	65.9 ± 8.79	.12†
Barrett’s esophagus length, cm		6.19 ± 4.18	4.0 ± 3.9	.07†
Hiatal hernia length, cm		2.9 ± 1.37	3. 4 ± 1.63	.27†
Body mass index, kg/m2		28.7 ± 5.43	30.62 ± 6.08	.27†
Tobacco use
Nonsmoker	14 (30.43)	7 (33.34)	7 (28)	.8‡
Ex-smoker	28 (60.87)	12 (57.14)	16 (64)
Current smoker	4 (8.70)	2 (9.52)	2 (8)
Alcohol use
No alcohol use	20 (43.48)	7 (33.33)	13 (52)	.115‡
Ex alcohol use	2 (4.35)	0 (0)	2 (8)
Current alcohol use	24 (52.17)	14 (66.67)	10 (40)
Follow-up (time between index EGD and development of neoplasia [in cases] or latest EGD [in control subjects]), y		3.7 ± 2.37	6.3 ± 3.1	.002†

Values are n (%), n or mean ± standard deviation.

^*Fisher exact test.

^†Welch t test.

^‡Kruskal-Wallis test.

As illustrated in Figure 1, histologic features are essentially identical in the index non-neoplastic biopsy samples from BE patients who developed dysplasia during follow-up (cases/progressors) and who did not (control subjects/nonprogressors). The biopsy samples revealed non-neoplastic specialized columnar mucosa with intestinal metaplasia, consistent with BE. The columnar epithelium contained basally located nuclei that were uniform in size, with smooth chromatin, indistinct nucleoli, and intact nuclear architecture. No cytologic or nuclear features are useful to predict which patient may develop dysplasia.

Figure 1.

Representative histology images of patients with Barrett’s esophagus (BE) who are (A, H & E, orig. mag. × 100) nonprogressors and (B, H & E, orig. mag. × 100) progressors. The histologic features are almost identical in the 2 biopsy specimens. Specialized columnar mucosa with intestinal metaplasia can be seen, consistent with BE, but there are no morphologic features to predict which patient may develop dysplasia.

nanoNAM-identified alterations in cases

nanoNAM measurements captured subtle submicron alterations in the nuclear architecture of the same case versus control as depicted in Figure 2. To comprehensively establish that nanoNAM-based characterization of epithelial cell nuclear architecture can indeed capture differential alteration between cases and control subjects, we generated drOPD cumulative distributions from the drOPD data cubes for the 2 patient groups. We focused on columnar-shaped nondysplastic epithelial cells only. Figure 3 shows that the drOPD cumulative distributions of the 2 cohorts did indeed undergo a differential shift. Moreover, using the Kolmogorov-Smirnov test,²¹ we found this differential shift was statistically significant with P < .001, suggesting the potential of nanoNAM to differentiate progressors from nonprogressors.

Figure 2.

Nanoscale nuclear architecture mapping (nanoNAM)-based visualization of histologically normal-appearing epithelial cells from unstained tissue sections from representative nonprogressor and progressor patients shown in Figure 1. A, Pseudo-colored tissue sample from a nonprogressor patient with Barrett’s esophagus (BE) with corresponding nanoNAM-based nuclear architecture of unstained epithelial cell nuclei, with zoom-in of 2 regions. B, Depicts the same for BE patient that undergoes neoplastic progression. The nanoscale sensitivity of nanoNAM-based characterization of nuclear architecture is able to capture the differences between progressors and nonprogressors architecture (redder values) that are not visible at the microscopic scale in Figure 1, with both samples appearing histologically normal.

Figure 3.

Cumulative distribution of nanoscale nuclear architecture mapping (nanoNAM) measurements. Depth-resolved optical path difference (drOPD)-based cumulative distributions of nanoNAM measurements showing a distinct and statistically significant difference between progressors and nonprogressors at the 95% confidence level with a P ≤ .03.

nanoNAM features quantify alterations in nuclear architecture from progressors to nonprogressors

We quantified the differences in the drOPD values of each patient using 12 nanoNAM features (Fig. 4). Of these 12 features, 6 (mean-maximum, mean-mean, mean-median, entropy-median, entropy-entropy, entropy-skewness) identified a statistically significant differential change in the nuclear architecture at a significance level of .05 between progressors and nonprogressors.

Figure 4.

Volumetric nanoscale nuclear architecture mapping (nanoNAM) summary features. Twelve nanoNAM features that summarize the intrinsic 3-dimensional (drOPD data cube) nuclear architecture of columnar-shaped nondysplastic epithelial cells from patients with nondysplastic Barrett’s esophagus who either underwent neoplastic transformation (progressors) or did not (nonprogressors). Their ability to a priori capture this progression is reflected by the statistical significance of their difference. ns, P > .05; *P < .05; **P < .01. drOPD, Depth-resolved optical path difference.

nanoNAM features predict progression of BE to neoplasia

Based on the above results, we trained 3 types of SVM classifiers to a priori identify those BE patients who progressed to HGD/EAC from those who did not. The 3 classifiers differed in the type of input. The first classifier used length of BE segment as the clinical input variable. The second used the 6 nanoNAM input features that show statistically significant difference between progressors and nonprogressors, whereas the third SVM combined both BE segment length and nanoNAM features as the classifier input. As shown by receiver operating characteristic curves plotted in Figure 5A, the nanoNAM features provided a significant improvement in performance over BE length alone. However, the combined nanoNAM and clinical features provided a marginal yet statistically significant improvement over nanoNAM features alone. This is demonstrated by the AUROC boxplots for the 3 SVM classifiers shown in Figure 5B corresponding to the 500 stratified bootstraps. SVM classifiers with BE length and nanoNAM features as separate input features have an AUROC curve of 67.96% ± .38% (mean ± standard error) and 80.8% ± .35%, respectively, with a statistically significant improvement in performance (P ≤ .0001). Importantly, combining BE length and nanoNAM features provided a marginal improvement in performance with an AUROC curve of 82.54% ± .46%, which is significant at the 99% confidence level with a P ≤ .01. Our results suggest that nanoNAM features–based SVM prediction of NDBE patients progressing to HGD/EAC improves with addition of BE segment length.

Figure 5.

Nanoscale nuclear architecture mapping (nanoNAM)-based prediction of neoplastic progression in patients with nondysplastic Barrett’s esophagus. A, Receiver operating characteristic (ROC) curve showing support vector machine (SVM)-based predictive performance, with 3 different input features: nanoNAM features, clinical variables, and nanoNAM+clinical features. B, Boxplots of area under the ROC curve for 500 bootstrapped training-validation pairs for the 3 separate input features. The statistical difference between the performance of the 3 SVM-based prediction models is visualized through the following notational convention: ns, P > .05; *P < .05; **P < .01; ***P < .001; ****P < .0001.

DISCUSSION

Despite being at an increased risk for developing EAC, most NDBE patients do not develop cancer. Current recommendations focus on the presence of dysplasia as a marker for neoplastic transformation in patients with BE. However, surveillance based on dysplasia can potentially be imprecise and suffer from interobserver variability. Our results suggest that nanoNAM has the potential to identify nuclear architectural abnormalities much earlier before the development of dysplasia and thereby help triage the small subset of BE patients truly at risk of EAC.

Several models based on clinical features, endoscopic findings, and molecular markers have been reported to predict the risk of progression in BE. A scoring system based on male sex, smoking, length of BE, and baseline LGD that identified patients with BE at low, intermediate, and high risk for HGD/EAC has been validated on an external cohort.^11,13 Another model based on a Swedish population registry incorporating age, male sex, and maximum BE length predicted 71% of all EAC/HGD cases; however, the study was limited because of missing data on endoscopic and clinical variables such as obesity and smoking.²²

Prior studies reported genomic and epigenetic changes in BE progressors. Aberrant p53 expression including both overexpression and loss of expression have been reported in patients at high risk for neoplastic progression.²³ In a meta-analysis of neoplastic risk with aberrant p53 expression, the odds ratio was 4 to 6 in case-control studies and 14 to 17 in cohort studies. However, the subjectivity in stain interpretation limits its widespread use. Tissue Cypher (Cernostics, Pittsburgh, Pa, USA) assay uses a multiplexed fluorescence imaging platform to assess nuclear morphology and 9 protein-based biomarkers (p16, alpha-methylacyl-coA racemase, p53, CD68, cyclooxygenase-2, CD45RO, hypoxia inducible factor-1α, HER2/neu, and cytokeratin-20). In a case-control study of 268 BE patients stratified by the prediction model based on Tissue Cypher, the high-risk group was at a 4.7-fold increased risk for HGD/EAC compared with the low-risk group (95% CI, 2.5-8.8; P < .0001).¹⁴ The sensitivity and specificity of the test at 5 years were 29% and 86%, respectively, with prevalence-adjusted positive and negative predictive values for risk of neoplastic progression of 23% and 96.4%, respectively.

In another study, genome-wide copy number instability assessed by shallow whole genome sequencing showed that genomic signals can distinguish progressors from non-progressors even 10 years before histopathologic transformation.²⁴ In yet another study, a prediction model including age, BE circumferential length, and a clonicity score over the genomic set including chromosomes 7, 17, 20q, and c-MYC resulted in an area under the curve of .88. The sensitivity and specificity of this model were .91 and .38, respectively, with positive and negative predictive values of .13 (95% CI, .09-.19) and .97 (95% CI, .93-.99).²⁵ This model can be implemented to identify NDBE patients who may be assigned to less frequent or no surveillance.²⁵

Several optical imaging techniques such as chromoendoscopy, autofluorescence imaging, confocal laser endomicroscopy, endocytoscopy, optical coherence tomography/volumetric laser endomicroscopy, and spectroscopy are used in enhanced imaging of BE; however, they are useful in the detection of concurrent dysplasia during endoscopy rather than risk stratification of BE patients for developing dysplasia. The nanoNAM approach uses the Fourier phase of the spectral interference between light back-reflected from within the cell nuclei and common-mode reference to interrogate the 3D structural alteration in the cell nuclei. The stability provided by the common-mode interference setup makes nanoNAM robust to phase noise, thereby allowing nanoNAM to detect submicroscopic abnormalities of nuclear architecture with nanoscale sensitivity of ~ 1 to 2 nm before their manifestation at the microscopic scale.¹⁸ The objective assessment of intrinsic nuclear architecture in NDBE is a particular strength of nanoNAM, because it does not use heuristics, empirically derived thresholds, or specific staining, avoiding inter- and intraobserver variability or staining variation. In addition, the nanoNAM approach is directly implemented on formalin-fixed paraffin-embedded tissue biopsy samples obtained as part of the standard clinical care without additional clinical or laboratory procedures. It could potentially be a cost-effective approach because it has minimal running cost and consumable reagent requirements and requires minimal training. Additionally, given its fast imaging speed requiring only about 40 minutes to image up to several hundred nuclei on a slide, together with data processing that can be performed on a standard computer, nanoNAM can seamlessly integrate with the existing clinical workflow. It is feasible to establish nanoNAM platforms in reference laboratories and tertiary medical centers with large specimen volumes. We therefore consider nanoNAM to be a quantitative extension of pathology to a submicron scale with a potential for clinical translation.

One limitation of our study is the relatively small number of patients in this cohort, and therefore we cannot make a categorical claim regarding integrated nanoNAM and segment length performance. However, this is a pilot study, and our data suggest that an integrated nanoNAM with clinical variables such as segment length could be an important facet of using nanoNAM in clinical settings. It is also not known whether spatial or temporal variations occur in the nanoNAM features in any given patient. In addition, the type of fixative and age of the specimen may potentially affect these nanoNAM features. Therefore, despite the strengths and the potential of nanoNAM, we are aware that a much larger study should be undertaken to expand and validate the results of this proof-of-concept study in the future.

Abbreviations:

3D

3-dimensional

AUROC

area under the receiver operating characteristic

BE

Barrett’s esophagus

drOPD

depth-resolved optical path difference

EAC

esophageal adenocarcinoma

HGD

high-grade dysplasia

LGD

low-grade dysplasia

nanoNAM

nanoscale nuclear architecture mapping

NDBE

nondysplastic Barrett’s esophagus

SVM

support vector machine