Potential Use of Quantitative Tissue Phenotype to Predict Malignant Risk for Oral Premalignant Lesions

The importance of early diagnosis in improving mortality and morbidity rates of oral squamous cell carcinoma (SCC) has long been recognized. However, a major challenge for early diagnosis is our limited ability to differentiate oral premalig-nant lesions (OPL) at high risk of progressing into invasive SCC from those at low risk. We investigated the potential of quantitative tissue phenotype (QTP), measured by high-resolution image analysis, to identify severe dysplasia/ carcinoma in situ (CIS; known to have an increased risk of progression) and to predict progression to cancer within hyperplasia or mild/moderate dysplasia. We generated a nuclear phenotype score (NPS), a combination of five nuclear morphometric features that best discriminate 4,027 ''normal'' nuclei (selected from 29 normal oral biopsies) from 4,298 ''abnormal'' nuclei (selected from 30 SCC biopsies). This NPS was then determined for a set of 69 OPLs. Severe dysplasia/CIS showed a significant increase in NPS compared with hyper-plasia or mild/moderate dysplasia. However, within the latter group, elevated NPS was strongly associated with the presence of high-risk loss of heterozygosity (LOH) patterns. There was a statistical difference between NPS of hyperplasia or mild/ moderate dysplasia that progressed to cancer and those that did not. Individuals with a high NPS had a 10-fold increase in relative risk of progression. In the multivariate Cox model, LOH and NPS together were the strongest predictors for cancer development. These data suggest that QTP could be used to identify lesions that require molecular evaluation and should be integrated with such approaches to facilitate the identification of hyperplasia or mild/moderate dysplasia OPLs at high risk of progression.


Introduction
Although the importance of early diagnosis in improving the mortality and morbidity of oral squamous cell carcinoma (SCC) has long been recognized, the disease is still frequently diagnosed late and prognosis has not changed for the last three decades (1).A major challenge for early diagnosis of the at-risk tissue is our limited ability to differentiate oral premalignant lesions (OPL) at high risk of progressing into invasive SCC from those at low risk.Oral cancer is believed to develop from OPLs, progressing from hyperplasia, through the increasing degrees of dysplasia to carcinoma in situ (CIS) and finally invasive SCC.Histology, the current gold standard, is reasonably effective in predicting malignant risk for severe dysplasia/CIS, of which 30% to 40% are likely to recur or progress even with aggressive surgical treatment (1,2).However, overall, it is a poor predictor for the majority of OPLs-those with hyperplasia or mild/moderate dysplasia, wherein most do not progress into cancer.The literature reports a wide range (11-36%) for the overall risk of malignant transformation, depending on the type of lesions being followed and the length of follow-up (3,4).Leukoplakia with dysplasia are more likely to progress to oral SCC than those without dysplasia (4); however, many studies have reported that for the individual lesion, the grade of dysplasia provides little indication of whether or not it will progress to cancer (4)(5)(6).
Alternative approaches that facilitate the identification of highrisk OPLs, particularly those within hyperplasia or mild/moderate dysplasia, need to be developed.The foremost development in such approaches over the last 5 to 10 years has been aimed at finding molecular markers (7)(8)(9).Among these, microsatellite analysis of loss of heterozygosity (LOH) has shown promise as a powerful adjunct tool for risk prediction (7,(10)(11)(12)(13)(14). Retrospective studies from our research team have shown that LOH pattern can classify morphologically indistinguishable hyperplasia or mild/moderate dysplasia OPLs into different risk categories.OPLs without LOH at 3p and 9p have the lowest risk of progression.In comparison those with LOH at 3p and/or 9p but not in other arms have a 3.8-fold increase in relative cancer risk.Those with LOH at 3p and/or 9p plus additional losses (at 4q, 8p, 11q, or 17p, high risk) have a 33-fold increase in relative cancer risk (14).
Given that the largest proportion of OPLs are hyperplasia or mild/moderate dysplasia, there is a need for an inexpensive system to triage such samples for molecular analysis.There is some indication that imaging devices that characterize phenotype changes associated with cancer risk could play such a role by complementing and adding to our current knowledge of risk assessment (15)(16)(17).Moreover, because exploration of phenotype provides a window to the functional status of the cell, showing whether the genetic alterations within are associated with structural/functional changes, this approach could expand our understanding of the biology underlying malignant progression.
Histopathologic evaluation of presence and degree of dysplasia is an appraisal of the tissue (and cell) phenotype using multiple variables.Quantitative tissue pathology (QTP) breaks down the components of this phenotype into multiple quantifiable units that can be studied independently and in combination, allowing the investigator to examine associations of such change with progression risk (Fig. 1A and B).For example, increased proportion of heterochromatin condensation is a high-risk nuclear characteristic for cancer (18).Traditional pathology judges this by one criterion, hyperchromatism, whereas QTP measures multiple chromatin features, such as whether the increased DNA is distributed around the edge of the nucleus or clustered in the center (Fig. 1B, left), whether the nucleus is dark with light areas or light with dark areas (Fig. 1B, middle), and whether the increased chromatin is evenly distributed (euchromatin) or clumped locally (heterochromatin; Fig. 1B, right).
In this paper, we present the preliminary results of an in-house QTP imaging system used to assess nuclear phenotype scores/ changes (NPS) in OPLs.This is the first study to correlate QTPdetected NPS with not only pathologic grading, but also molecular changes and outcome of OPLs.The results showed that NPS correlates strongly with histology grading, genetic damage (LOH), and cancer progression.

Materials and Methods
Sample source.A total of 128 oral premalignant biopsies were collected from 128 subjects.All these lesions were obtained from the British Columbia Oral Biopsy Service with selection based on the criteria below.These lesions were grouped into a training set and a test set.
The training set contained two groups of specimens representing the two extreme ends of the intraepithelium spectrum: 30 oral biopsies with areas of relatively normal oral mucosa (e.g., amalgam tattoo or melanotic macule) and a group of 29 SCC.Of the 29 SCC, the majority were from the high-risk oral site ( floor of mouth and lateroventral tongue, 66%), were well to moderately differentiated (83%), and were stage 1 or stage 2 tumors (69%).
The test set contained 69 biopsies that can be separated into two subgroups: a group of 25 biopsies diagnosed as severe or CIS which was denoted as high-histology risk group because they are histologically at risk for progression (1, 2) and a group of 44 biopsies which were a subset of 83 well-characterized OPLs with sufficient material for QTP used in our previous study that explored the use of LOH to predict risk of progression for OPLs with hyperplasia or mild/moderate dysplasia (14).These 44 lesions (one biopsy per patient) diagnosed as hyperplasia, mild or moderate dysplasia are denoted the low-histology risk group.They were from patients without a prior history of head and neck cancer.The only prerequisite for inclusion in the present study was the availability of two adjacent unstained slides that were serial to slides used for the previous LOH analysis.Of these lesions, 15 have progressed to cancer and the remaining has not.Among the 15 progressing lesions, there were six hyperplasia, five mild dysplasia, and four moderate dysplasia.Among the nonprogressing lesions, there were 15 hyperplasia, 8 mild dysplasia, and 6 moderate dysplasia.There was no difference between the progressing and the nonprogressing lesions in terms of gender, age distribution, and smoking history (all P < 0.05).On average, the nonprogressing cases have been monitored for over twice the duration of progressing cases (71 mo versus 30 mo) to ensure that progression did not occur.
Sample preparation.All 128 samples were formalin fixed and paraffin embedded.The histologic diagnoses of the samples were confirmed by two oral pathologists (C.P. and L.Z.).Serial sections, 4-Am in thickness, were cut from each sample and placed on two glass slides; one was stained with H&E and the other with Feulgen-Thionin (19).The area on the H&E slide selected by the pathologist (C.P. or L.Z.) for conventional histopathologic diagnosis was the exact same area that was circled on the Feulgen-stained slide to be examined with the QTP imaging system.
QTP imaging system.This system is a modified version of the Cyto-Savant automated quantitative system (Cancer Imaging, BC Cancer Agency; ref. 19).The illumination wavelength was 600 F 5 nm, corresponding to the absorption peak of the Thionin stain.The effective pixel sampling space within the plane of the sample was 0.34 Am 2 , and the effective pixel sampling area was 0.116 Am 2 .The software is specifically designed for interactive semiautomatic cellular and architectural analysis of tissue sections (19,20).A strict quality control procedure was implemented to ensure the stability of the imaging system for each analysis (21).The imaging system characteristics followed the recommendation of the European Society of Analytical Cellular Pathology (22).
QTP image analysis.An experienced cytotechnician under the guidance of an oral pathologist (C.P. or L.Z.) delineated the region of interest (ROI) from the Feulgen-stained slide under a 20Â objective.The area selected represented the most abnormal tissue according to the pathologist.The analysis of collected nuclei in each biopsy was done automatically by the system and consisted of four steps: (a) delineation and focusing of the ROI (Fig. 1A), (b) automatic thresholding and segmentation of the nuclei in the field (23), (c) interactive correction of segmentation errors and selection of only nonoverlapping cells, and (d) automatic collection of individually focused images of each selected cell (Fig. 1D).For normal epithelium and OPLs, the full width of the epithelium had to be present.Only in-focus, intact nuclei were collected by the cytotechnician with nuclei obtained from all layers of the epithelium.Basal and parabasal cells may be preferentially chosen since there are more of them, but they are also more likely to overlap or be in clusters (higher density) and so will be rejected from the analysis as being overlapping cells.Intermediate and superficial cells, although less numerous, are more often isolated, correctly segmented, and thus more often selected for analysis.
The regions used in each case of SCC were selected using three criteria: (a) the region of tumor selected needed to show tumor differentiation that was representative of the case; (b) the region had to have minimal amounts of inflammation and tissue; and (c) the tumor region needed to be reasonably large, i.e., not a single line or layer of tumor cells.
On average, >100 cell nuclei were collected from the delineated area for each biopsy.For each nucleus, 110 nuclear phenotype features were calculated by the system, measuring the size, the shape, the amount, and the distribution of stained DNA in the nucleus (24).
Calculation of NPS.To determine nuclear features that are predictors for OPLs with different cancer risk, we first trained the imaging system to separate normal specimens (30 samples) from cancer specimens (29 SCC samples) in the training set.This was done by establishing first a cell-by-cell NPS, which is a linear discriminant function of the nuclear features that are statistically significant, not collinear, that taken together, used as a predictor, can classify cells into separate categories.For this process, all epithelial cells (4,027 cells) from the 30 normal samples were grouped together and used to define the ''normal-like'' cell group, and all epithelial cells (4,298 cells) from the 29 SCC samples were grouped together to define the ''cancer-like'' cell group.A forward stepwise discriminant function analysis of the 110 nuclear phenotypic features was done to select which of these features could be used to generate a discriminant function that best separated the two groups of cells.All features were considered as continuous variables.We used the default values provided by STATISTICA package: tolerance value, 0.01; F to enter, 1.0; F to remove, 0.0.In stepwise discriminant function analysis, a model of discrimination is built step-bystep, automatically done by the software.Specifically, at each step, all variables are reviewed and evaluated to determine which one will contribute most to the discrimination between groups.That variable will then be included in the model, and the process repeats.The process stops when the resulting correct classification rate (percentage of cells correctly classified in each group) no longer increased with additional variable.
An interpretation of this discriminant function is that it represents a cellby-cell phenotype classification procedure, which assigns a score to each nucleus indicating similarity to normal or cancer cells.
Similar to the method described in a companion paper (15), the range of this cellular score was divided into 10 continuous regions: the first region A represents values of phenotypically normal-like cells, and region J represents values of phenotypically cancer-like cells.
From the percentage of cells in each of these 10 regions, the NPS assigned to each biopsy was calculated as NPS = 1*A + 2*B + 3*C + 4*D + 5*E + 6*F + 7*G + 8*H + 9*I + 10*J, wherein A is the proportion of cells whose disciminant score falls in the region A, B is the proportion of cells whose disciminant score falls in the region B, and so forth.
This NPS represents the weighed sum of the 10 regions.The possible values for the NPS range from 1 to 10.An NPS with a value of 1 corresponds to a specimen containing 100% of normal-like cells, whereas an NPS with a value of 10 corresponds to a specimen containing 100% of cancer-like cells.
Statistical analysis.For comparison of difference between groups, a nonparametric Mann-Whitney U test was done.For comparison of difference between more than two groups, a nonparametric Kruskal-Wallis ANOVA was done.When the test showed a statistical difference across

Cancer Research
Cancer Res 2008; 68: (9).May 1, 2008 groups, a multiple comparison of mean ranks was done to confirm significant differences across group medians (multiple comparisons of mean ranks for all groups).Differences between group frequencies were assessed with the Yates corrected m 2 test.Time-to-progression curves were estimated by the Kaplan-Meier method, and comparisons were done using the log-rank test.Relative risks were determined using Cox regression analysis.All analyses were done with Statistica6 (StatSoft, Inc.).All P values were two-sided.A two-sided P value of <0.05 was considered statistically significant.

Results
NPS.We used a forward stepwise discriminant function analysis that selected 5 of the 110 nuclear features that best discriminate cells from normal and cancer samples in the training set: (a) Max_radius, the maximum radius of the nucleus; (b) Harm003_ fft, the third order harmonic of the nuclear boundarymeasures how much of the nuclear boundary can be explained by three lobes; (c) Fractal_area1, the relative spatial distribution of high and low absorbance variations in the nucleus-a measurement of heterochromatin versus euchromatin organization; (d) absorbance_skewness, measures whether the nucleus is dark with light areas or light with dark areas; and (e) Long90_runs, a measurement of the fraction of the nuclear diameter one can travel before an intensity change is encountered (24).A total of 94% of the 4027 normal-like cells and 77% of the 4298 cancer-like cells were correctly classified using these five discriminant features.These features were used to generate the cell-by-cell discriminant scores which were amalgamated across all of the cells selected in the epithelium to generate the NPS for each specimen.Figure 2 shows representative images of the ROI of four oral mucosa specimens  Correlation of NPS with pathology.Figure 3 shows the association between NPS and the pathology diagnoses.There was a monotonic increase in NPS with severity of pathology diagnosis [NPS median (range)]: 3.7 (2.01-5.03)for normal; 4.14 (3.03-7.90)for the low-histology risk group, i.e., the hyperplasia or mild/ moderate dysplasia group; 5.24 (3.65-8.03)for high-histology risk group (severe dysplasia or CIS); and 7.25 (3.77-9.96)for SCC.There is a significant increase in NPS between the low-histology risk group (hyperplasia or mild/moderate dysplasia) and the highhistology risk group (P = 0.036).There are significant differences between all logical histologic groupings except normal versus hyperplasia or mild/moderate dysplasia and severe dysplasia/CIS.The increase in NPS between the severe dysplasia/CIS group and the SCC was close to significance (P = 0.08).
The remaining analyses were done using only the lesions classified as hyperplasia, mild dysplasia, and moderate dysplasia (hyperplasia or mild/moderate dysplasia group).
Correlation of NPS with progression to cancer.We sought to determine whether NPS would identify hyperplasia or mild/ moderate dysplasia at high risk of developing into invasive SCC by comparing NPS values in progressing and nonprogressing lesions.Progressing cases showed significantly higher NPS value compared with nonprogressing lesions (NPS median, 5.7 for progressing lesions versus 3.8 for nonprogressing lesions; P < 0.0001).As shown in Fig. 4A, although the sample set was small, this trend remained statistically significant when the hyperplasia and dysplasia were examined separately [NPS median, 4.9 for progressing hyperplasia versus 3.8 for nonprogressing hyperplasia (P = 0.016) and 6.3 for progressing low-grade dysplasia versus 3.8 for nonprogressing low-grade dysplasia (P = 0.001)].
Cutoff NPS value and time to progression to cancer.For the NPS to be used by clinicians in cancer prediction, a cutoff value is needed above which clinicians could expect increased cancer risk for an OPL.We have chosen a cutoff value of 4.5 for the NPS because it provided the best separation (area under the ROI curve; Fig. 4B) between the nonprogressing lesions and the progressing lesions.In the training set (normal and SCC; Fig. 2), only 5 of the 30 normal samples (17%) had an NPS value of z4.5 (high NPS) whereas 28 of the 29 SCCs (97%) had high NPS.
When the hyperplasia or mild/moderate dysplasia lesions were categorized by this cutoff value, 17 of the 44 hyperplasia or mild/ moderate dysplasia test study cases had high NPS.Thirteen of the 17 cases (76%) with high NPS progressed into invasive SCC in contrast to only 2 of the 27 cases (7%) with low NPS (P < 0.0001).In other words, with this cutoff, we could correctly classify 86% of nonprogressing cases (25 of 29) and 86% of progressing cases (13 of 15).Time-to-progression curves were plotted as a function of the NPS classification into one of the two groups (high NPS versus low NPS; Fig. 4C).A significant difference was observed between the low-NPS group and the high-NPS group (P < 0.00015).
The proportion of lesions progressing to cancer at 5 years was 71% (12 of 17) for lesions with high-NPS and only 22% for lesions with low NPS.There was a 10-fold (relative risk, 10.3; confidence interval, 2.9-60.0)increase in the relative risk of progression to cancer for oral lesions with high NPS compared with those with a low NPS.
Correlation of NPS with LOH.Table 1 shows association between NPS and different LOH patterns, indicating the proportion of cases within each LOH pattern that are above and below the NPS cutoff.NPS values were consistently higher for samples with LOH at all seven chromosome regions examined.This increase was significant for 3p (P = 0.03), 4q (P = 0.003), 9p (P = 0.003), and 11q (P = 0.05).We also looked at associations with multiple losses and for LOH at 3p and/or 9p plus LOH at any of the arms 4q, 8p, 11q, 13q, and 17p, patterns previously identified as having markedly increased risk of progression (9,11).NPS was strongly associated with the presence of these high-risk LOH patterns: multiple losses (P = 0.0005) and LOH at 3p and/or 9p plus LOH at any of the arms 4q, 8p, 11q, 13q and 17p (P < 0.0005).
Seventeen of the 44 hyperplasia or mild/moderate dysplasia lesions had high NPS.Molecularly, 16 of these 17 cases had LOH information.Of these 16, 11 (69%) showed LOH at 3p and/or 9p, a loss believed to be essential although not sufficient for cancer progression (8); in contrast, only 7 of the 27 cases (26%) with NPS of <4.5 (low NPS) show such a loss (P = 0.008).Similarly, 9 of these 16 lesions (56%) showed LOH at 3p and/or 9p and at any of the other five chromosomes, which corresponds to the highest risk pattern for cancer progression (14), whereas only 4 of the 27 cases (15%) with NPS of <4.5 (low NPS) showed such a pattern (P = 0.007).
Regression analysis with multiple covariates.We did a regression analysis with multiple covariates for the hyperplasia or mild/moderate dysplasia group, looking at histology, LOH, and NPS with cancer progression as primary end point.For each of these covariates, we assigned a value of either 0 or 1, denoting either lowrisk or high-risk marker value, respectively.For histology, hyperplasia was coded 0 and mild or moderate dysplasia was coded 1.For LOH, cases with LOH at 3p and/or 9p plus LOH at any of the arms 4q, 8p, 11q, 13q, and 17p were coded 1 and 0 if otherwise.For NPS, cases with high NPS (NPS, >4.5) were coded 1 and cases with low NPS were coded 0. The results of the Cox proportional hazard regression are given in Table 2.In this table, the first three models examine NPS, LOH, and histology separately and show that only LOH and NPS are significant in predicting cancer (with similar P value, P < 0.001).For hyperplasia or mild/ moderate dysplasia, histology is not a predictor of cancer progression (P = 0.88).Categorization of hyperplasia with mild dysplasia versus moderate dysplasia was also not significant (data not shown).When LOH and NPS were entered together into the Cox regression analysis (model 4), NPS is no longer as significant as it was when used as a single predictor (model 1, P = 0.0007), although the borderline P value of 0.07 does suggest that a larger sample size may be required for a definitive conclusion.The different models clearly show that conventional histology has no predictive effect and that NPS is a better ''phenotypic'' predictive marker for cancer risk than histology (models 5 and 6).

Discussion
There have been tremendous scientific advancements in many fields over the last decade, owing a significant part to rapid computer technology development; such development is revolutionizing many areas, including the health care system.In this study, we explored the value of a computer-driven imaging system as an adjunct tool to assist the pathologist in judging the progression risk of OPLs.The importance of dysplasia phenotypes as cancer risk predictors is well recognized.The strong association of marked dysplasia with increased risk of cancer progression has been observed and confirmed in multiple sites (e.g., lung, esophagus, breast, cervix, and skin) and is the reason why histopathologic interpretation of dysplasia is the current gold standard.However, this standard is far from accurate because of a number of issues: (a) by nature it is subjective, and studies have reported low intraobserver and interobserver agreement in grading epithelial dysplasia (25,26); (b) there is a lack of knowledge of the weight to be associated with each of the visual characteristics used to assess histopathologic grades, as well as the way in which they interact for risk assessment; and (c) reactive changes (with little malignant potential) can cause alterations that resemble low-grade dysplasia, making differential diagnosis in early disease particularly challenging.Finding a solution to the above issues involves at least two parts: (a) the development of a tool that can be made to be high throughput, that is objective, quantitative, and sensitive enough to unravel subtle differences in OPLs, such as between reactive and true low-grade premalignant changes and between nonprogressing and progressing low-grade dysplasia and (b) the collection of specimens with known risk factors, including highgrade histology and high-risk molecular pattern, and ideally with outcome to test the system.
Our results showed that the in-house imaging system developed by our research team, although not yet fully automated and high throughput, is quantitative and can unravel subtle differences in OPLs.We used two well-defined groups of cells (normal and cancer) to generate an NPS for each specimen.This system was able to analyze 110 nuclear features and identify not only differences in nuclear size and chromatin density as noted by pathologists, but also differences in nuclear shape and chromatin texture, features not readily quantified by pathologists.The identification of different features of chromatin texture changes by the system is significant because such changes are now widely accepted as an indication of genetic or epigenetic changes (18).Furthermore, the system also quantified the changes measured, allowing an objective computation of an NPS for each sample.
In a previous study (15), we determined how sensitive the NPS process is to the interactive steps involved in the collection of the  4), 4.6 (5), 4.7 (6), 4.8 (7), and 4.9 (8).C, probability of having no progression to cancer for specimens with low NPS (NPS, <4.5) and specimens with high NPS (NPS, >4.5).
nuclei used to calculate NPS.In that study, 30 preneoplastic bronchial biopsies which spanned the ranges of pathology grades and NPS were selected and measured.All samples were measured twice by one experienced cytotechnologist and twice by another experienced cytotechnologist.The measurement of the NPS was highly reproducible between individual technologists.The coefficient of correlation between two histopathology technologists was 0.98.The intratechnologist variability was comparable.
The NPS was noted to continuously increase with severity of histologic grading (P < 0.0001).A significant difference was observed between the low-histology risk hyperplasia or mild/ moderate dysplasia group and the high-histology risk group (severe dysplasia/CIS).In British Columbia, a histologic diagnosis of severe dysplasia/CIS results in the majority of cases receiving treatment by surgery.Thus, we were unable to examine a potential association of NPS in progression within this category.However, 72% of the severe dysplasia/CIS in this study had NPS of >4.5 indicating a strong association of NPS with this high-risk histologic category.
This is the first study to investigate the relationship between NPS and outcome for hyperplasia or mild/moderate dysplasia OPLs, which is ultimately the most important aspect of validation of any new test.High NPS was found to be strongly associated with cancer progression (P < 0.0001).This association was independent of histology because high NPS in oral lesions with low-grade or no dysplasia was still strongly associated with cancer progression, although the number of cases were small.These results were also assessed by multivariate Cox regression analysis and showed an absence of predictive value for histology classification among hyperplasia or mild/moderate dysplasia lesions.In a mixed model, LOH is still the most significant predictor, but a larger sample size is required to determine whether QTP adds significantly to the correlation in a mixed model.It does support the use of QTP for triaging samples that require LOH analysis.This is the first study to investigate the relationship between NPS and molecular changes of premalignant lesions.We used the microsatellite analysis for LOH, a technique currently used by many research groups in the study of cancer risk of OPLs.Increased NPS was significantly associated with the presence of previously reported high-risk LOH patterns, including the presence of multiple losses and LOH at 3p and/or 9p plus additional losses (at 4q, 8p, 11q, or 17p; refs.[11][12][13][14]. We have determined an NPS cutoff value of 4.5 based on the training set (normal and SCC; Fig. 2).Seventeen of the 34 hyperplasia or mild/moderate dysplasia cases showed high NPS.Sixteen of these 17 cases had LOH data available, and all showed LOH at 3p and/or 9p, a loss believed to be essential although not sufficient for cancer progression (14).Eleven of these 16 cases progressed into invasive SCC.Individuals with high NPS have a 10-fold increase in relative risk of developing a cancer within 5 years compared with individuals with a low NPS (confidence interval, 2.05-46.6).Similar results were obtained when the NPS threshold used to define the high NPS and the low NPS groups was varied.NPS could therefore offer clinicians additional information on the possible malignant potential of lesions.For example, 6 of the 21 hyperplasia had high NPS, five of these six had a molecular pattern available, and four of them showed the high-risk LOH pattern.All three of the hyperplasia that progressed into cancer had these characteristics.Of the 23 low-grade dysplasia, 11 (48%) had high NPS and 9 of these had LOH at 3p and/or 9p (7 had LOH at 3p and/ or 9p plus another loss).Eight of the 11 progressed into cancer.These results suggest that the QTP image system could be a powerful adjunct tool and reproducible method to identify hyperplasia or mild/moderate dysplasia lesions for molecular assessment.
These findings are similar to results obtained by our team in bronchial preneoplastic lesions (15).Using the same approach as the one described in this paper, we showed that QTP correlates better with genetic damage and cancer progression than conventional pathology.Bronchial NPS is currently being used as a surrogate end-point biomarker in a chemoprevention trial in bronchial dysplasia (27).Similarly, using different morphometric features, quantitative analysis of endometrial hyperplasia correlates better with genetic clonality than conventional pathology (28).Similar results by other groups have been found in bladder, prostate, cervix, and breast cancer for the association of QTP with histologic grading (29)(30)(31).Regarding the nuclear features used in the calculation of our NPS, they are similar to the features used in other studies for the calculation of morphometric scores, such as quantitative nuclear grade by Veltri (29) in transitional cell carcinoma, discriminant score by Hanselaar et al. (32) in cervical cytologic specimens, and morphometric nuclear grade by Bacus et al. (16) in breast and cervical specimens.The main consistency between all these studies, different in terms of tissue types and specimens (cytology/histology), is that they all used a combination of different types of features, i.e., features describing the nuclear shape and size, as well as features measuring chromatin texture.Furthermore, nuclear feature calculations can easily be reproduced as described in Rodenacker (33).
Recently, Montinori et al. (17) has shown that these findings could be reproduced using H&E-stained material and fully automated analysis to predict recurrence in urothelial lesions.If this work is confirmed, it could be a definitive breakthrough for the implementation of quantitative histology in routine histopathology, as both conventional and quantitative analysis would be done on the same material, with the analysis fast and automated.
A large longitudinal study is needed to confirm the prognostic value of QTP, and such a longitudinal study is under way.The overall goal of that study is to evaluate three new innovative, potentially complementary devices for early detection and followup of OPLs and oral cancer and to determine their potential role in the triage of patients in a provincial network that is being established in British Columbia (34).One device is an in vivo clinical tissue evaluation tool; the others involve the development of quantitative pathology approaches.
In conclusion, the results presented in this paper show that NPS correlates strongly with genetic damage, cancer progression, and histology grading and could have a potential value for triaging OPL for molecular analysis.

Figure 1 .
Figure 1.QTP of an oral dysplastic epithelium.A, ROI is manually delineated according to pathologist diagnosis.Nuclei selected by the technician within the ROI are automatically segmented (blue ).B, graphic representations of 3 of the 110 nuclear features assessed in this study, all of which measure DNA distribution in the nucleus.Left, OD-Skewness measures whether the nucleus is dark with light areas or light with dark areas; middle, Fractal_area1 measures heterochromatin versus euchromatin organization, i.e., large intensity contrast between highly condensed chromatin and noncondensed chromatin; right, Long90_Run measures the fraction of nuclear diameter one can travel before an intensity change is encountered.C, histogram distribution of the DNA content, 1 of the 110 nuclear features.D, gray level images of a subset of selected nuclei.
with the corresponding histogram distribution of texture feature Fractal_area1 used in the calculation of the NPS.To determine the ability of NPS to classify oral lesions into different risk groups, we correlated NPS with (a) pathology classification, (b) molecular patterns, and (c) outcome, the ultimate yardstick for judging the validity of new diagnostic tools.

Figure 2 .
Figure 2. Images of the ROI of four oral mucosa specimens with the corresponding histogram distribution of texture feature Fractal_area1 used in the calculation of the NPS.First row, normal epithelium (top ); second row, nonprogressing mild dysplasia; third row, progressing mild dysplasia; fourth row, SCC.

Table 2 .
Analysis of cancer risk by combining NPS, LOH and Histology using the Cox proportional regression hazards model NOTE: Models 1 to 3 correspond to univariate models.Models 4 and 5 represent models with combination of markers.

Table 1 .
Proportions of lesions with low and high NPS medians in groups with different LOH patterns