Running Title: Cell Image Analysis for Esophageal Dysplasia

Bin Zhou, Uta Jütting, Karsten Rodenacker, Peter Gais, Pei-Zhong Lin

Bin Zhou is cytopathologist, guest scientist of Cancer Institute (Hospital), Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China

Uta Jütting and Karsten Rodenacker are Diplom Mathematicians, staff members and senior scientists, GSF-National Research Center for Environment and Health, Institute for Biomathematics and Biometry, D-85764 Oberschleißheim, Germany

Peter Gais is Diplom Engineer, staff member and senior scientist, GSF-National Research Center for Environment and Health, Institute of Pathology, D-85764 Oberschleißheim, Germany
Pei-Zhong Lin is Prof. of Cytopathology, Cancer Institute (Hospital), Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China

Address of correspondence: Uta Jütting, GSF-Institute for Biomathematics and Biometry, Ingolstädter Landstr. 1, D-85764 Oberschleißheim, Germany. Tel.: +49 89 3187 2555; Fax: +49 89 3187 3127, e-mail: uta.juetting@gsf.de.

ABSTRACT

Objective: The principal reason for the high resolution image analysis of esophageal dysplasia is that of biomarkers of precancerous lesions which are urgently needed as endpoints in the field of chemoprevention. Previous studies have shown that there are around 25% esophageal epithelial dysplasia in mass screening from 45-69 year old persons. But there are about one third esophageal dysplasia patients develop to carcinoma within several years. Some cases with dysplasia are static and some regressing to normal. It is unable to discriminate one from the other by cytopathologist under the microscope particularly by discrimination of the nonspecific cell alternations and precancerous lesions from the invasive. This paper tries to evaluate the way of surrogate endpoints biomarkers for cancer incidence in chemoprevention trials.

Study Design: Asymptomatic adults were examined with balloon sampler in 1983 from Heshun Commune of Linxian County. 50 cases of esophageal moderate dysplasia and 68 cases of esophageal mild dysplasia were selected for this study. By means of an Axiomat-microscope equipped with a TV-camera 100 visually normal intermediate squamous cell nuclei per specimen were randomly measured from routinely Papanicoalou stained slides.

Results: Of 50 esophageal moderate dysplasia cases 24 and 7 progressed to carcinoma within 3 and 9 years resp.. The other 19 cases remained stable or regressed to nomal and were used as control group. By means of chromatin features correct specimen classification rates of 79.2% (19/24) and 73.7% (14/19) and 85.5% (6/7) and 84.2% (16/19) were achieved, respectively (p<0.001).

Of 68 cases classified as mild dysplasia 16, 13 and 12 cases progressed to carcinoma within three, five and nine years, resp.. The other 27 cases remained stable or regressed to normal and were used as control group. The correct specimen classification rates were 93.8% (15/16) and 88.9% (24/27), 69.2% (9/13) and 74.1% (20/27), and 83.3% (10/12) and 77.8% (21/27) resp, using chromatin features of the nuclei (p<0.001).

Conclusions: In this study chromatin nuclear features measured by high resolution image analysis can sufficiently good forecast the outcome of precancerous lesions and discriminate precancerous lesions with progression and non-progression. It also can be employed as surrogate endpoint biomarkers in clinical chemoprevention trial. Stoechiometric staining and standard preparations should increase the correct classification rates in our further studies.

Keywords: high resolution image analysis, esophageal dysplasia, surrogate endpoint biomarkers, chromatin features, malignancy associated changes, MAC

INTRODUCTION

Esophageal cancer is a common disease with very poor prognosis. Nearly 50 percent of the world's esophageal cancer occurs in China (22), where it is the second leading cause of cancer death after stomach carcinoma (16). The Taihang mountain area of north-central China has recorded some of the highest esophageal cancer mortality rates in the world. In Linxian, a high-risk county in Henan province, the age-adjusted mortality rates for esophageal cancer in 1973-1975 were 161 per 100,000 for men and 103 per 100,000 for women, and at age 75 there was a cumulative mortality from esophageal cancer of over 20% for both sexes (15,16). Esophageal cancer has been an important cause of death in this area for hundreds of years.

Over the past 40 years, Chinese scientists have developed esophageal balloon cytology (EBC) as an early detection technique for identifying early cancerous lesions and for screening asymptomatic precancerous lesions in high-risk populations (28, 29). Since that time, systematic research work has been done on precancerous lesions (also called intraepithelial neoplasia or dysplasia) of the esophagus. Previous studies have reported that dysplasia of esophageal epithelium is an essential step of esophageal carcinogenesis. They are already validated as predictors of cancer incidence (17). The conclusions suggest that, for the prevention of esophageal carcinoma it is essential and possible to treat the precancerous lesion by blocking its progression and promoting its regression to normal (17). A large scale chemoprevention of esophageal carcinoma study was launched in 1983. This trial enrolled 2531 men and women, between 40 and 65 years old, who had no previous history of cancer but cytologic evidence of esophageal dysplasia. The clinical chemoprevention trial began in 1983 and ended in 1993. It was designed as a 2-arm randomized placebo-controlled chemoprevention trial in which one half of the participants received daily pills containing Anti Tumour B (ATB, a pure Chinese traditional medicine) or Retinamide and the other placebo pills. After 3 , 5 and 9 years, all living participants were invited to have again a EBC procedure. More than 90% accepted. Several months after these EBC exams, all individuals with a EBC diagnosis of severe dysplasia were also invited to undergo endoscope. The results indicated that ATB and Retinamide have beneficial effects in reducing the progression of the esophageal carcinognesis process and reduced the incidence rate of esophageal carcinoma around 50% (p < 0.01), (18, 19).

As it is well known that esophageal epithelial dysplasia is an unsteady state, either developing to malignancy or regressing to normal. In fact, only a very small proportion of dysplasia will progress to cancer. Therefore, dysplasia is characterized by its capability to have two transformation pathways (17), however, even today, most of the morphological methods have a low degree of diagnostic specificity, somewhat subjective, and it is nearly impossible to get information on deeper layers of the cells. This is partially a result of difficulties in discriminating the cells of different lesions and of different biological behavior using morphological analogy, particularly of non-specific cell alterrations and precancerous lesions from invasive areas. It is a serious barrier to development of the field of chemoprevention as biomarkers. The biomarkers of precancerous lesions are urgently needed as endpoints in clinical trials of chemoprevention agents because they require less time, cost and effort compared to the conventional endpoint of cancer incidence reduction. It is highly desirable that early, precancerous lesions are identified which can serve as surrogate endpoints biomarkers (SEB) for cancer incidence in chemoprevention trials and, more importantly, as targets for chemoprevention (10). Obviously, the human cannot transform the visual impression of a nucleus into anything similar resembling a histogram, thus genuinely new information has to be gained for the diagnostic process. The high resolution image analysis may be able to appreciate very slight but consistent differences in nuclear appearance changes that are too subtle for a cytologist to observe.

MATERIALS AND METHODS

Subjects and Specimens:

Participants were recruited in the spring of 1983 from Heshun Commune of Linxian County, which is well-known for its high rate of esophageal cancer. All of the 40-65 year olds were asked to participate, unless they had a history of cirrhosis, esophageal varices, vomiting blood, or were considered too weak to undergo the examinations. 50 cases of esophageal moderate dysplasia (a Chinese cytologic diagnosis of dysplasia I and II) and 68 cases of esophageal mild dysplasia were selected for high resolution image analysis (Table 1). During the next three and nine years, there were 24 and 7 cases of moderate dysplasia progressed to esophageal carcinoma. The other 19 cases remained stable or regressed to normal. There were 16, 13 and 12 cases of mild dysplasia progressed to esophageal carcinoma the next three, five and nine years, resp. The other 20 cases remained stable or regressed to normal. The diagnoses, the given class names, the number of specimens and cells measured of each group are also listed in Table 1. All individuals were followed up prospectively, with monthly visits by village doctors, cytologic and endoscopic examinations of symptomatic subjects, and additional cytologic screening examinations of selected subjects in 1987 and 1993. Case records and diagnostic materials (cytologic slides, histologic slides and x-rays) were reviewed and the cancer diagnoses were confirmed by cytopathologists, pathologists and radiologists from Cancer Institute, Chinese Academy of Medical Sciences of China.

The balloon sampler used in this study was a rubber balloon covered with a cotton mesh net and attached to a double-lumen rubber tube. This balloon sampler has been used in most of the mass population screenings performed in China during the last 40 years. After balloon technique, the slides were immediately fixed in 95% ethanol for fifteen minutes, and later stained with standard Papanicolaou stain.

Data acquisition:

By means of an Axiomat-microscope (Zeiss, Oberkochen, Germany), equipped with a TV-camera (Bosch, T1VK9B1, Stuttgart, Germany, 128x128 pixel) 100 visually normal well preserved intermediate squamous cell nuclei per specimen were randomly selected and digitized. The nuclei were scanned in transmission with a 100x objective (oil immersion, numerical aperture 1.3) using an optical narrow band filter of 548 nm wavelength. The pixel distance was 0.25 µm, and the nominal grey value resolution was covered by 256 channels. Image processing was carried out using a VAX 4000-500 processor (Digital, Maynard, USA) with software written under idl (Interactive Data Language, RSI, Boulder, Colorado, USA). Each nucleus was automatically segmented, visually controlled and interactively improved if necessary. More than 100 quantitative features (morphological, densitometrical, textural) were extracted using the extinction or optical density image, which was derived from transmission image (24). A shading correction was applied. Using linear and non-linear filtering like Roberts gradient, Laplace transform, flat texture image, local fractal and multi-fractal dimensions, topological gradient, the difference of upper and lower skeleton, and statistical features derived from runlength and co-occurrence matrix (23, 27) several chromatin distribution features were calculated from the whole nucleus as well as from dark and bright regions of the nucleus. The latter were automatically discriminated using the upper and lower skeleton which is similar to the watershed algorithm applied to the extinction image and its inversion (upper and lower skeleton). Due to nonstoechiometric staining and data acquisition over years only those features have been offered for classification that were proved to be nearly independent of staining intensity, in order to avoid variances due to preparation and staining influences (25). The remaining feature set contained about 60 variables. In the appendix the finally selected features are listed and shortly described.

Statistics:

All statistical evaluations were done using SAS (SAS Institute, Inc., Cary, North Carolina, USA) and BMDP (Statistical Software Inc., Los Angeles, CA, USA) program packages. All cells from specimens belonging to the same clinical sample were pooled and two-class stepwise linear discriminate analyses on cell level were applied. From the whole feature set only those features were offered in the classification steps which are univariate significant and not highly dependent from staining and preparation changes. Up to 10 features were stepwise selected on the basis of F-statistics. The value for the first chosen feature is the univariate one whereas the following F-values are multivariate reflecting the impact of results after using this feature together with the already selected features. For each specimen, the mean of the a posteriori probability (APOP) distribution of the corresponding cells was calculated. The APOP value and the double standard error of this mean (S.E.M.) were used for specimen classification. A specimen was classified into that class with the highest APOP value only if the mean APOP±S.E.M. did not cut a threshold (THR) which was set as the border between the two classes. In all cases, the threshold was defined at APOP=0.5 that is the half distance between both group means. Cases with THR{APOP±2S.E.M.} were called unclear (5). The significance for the specimen classifications was tested using contingency tables without defining unclear cases. In these cases and in the three-class discrimination the specimens were classified into that class with the highest APOP value. All statistical evaluations were done at 95% level.

Results

All investigations in the following concentrate on the discrimination of cell nuclei from different defined dysplasia classes and their progression status.

- Discrimination of patients with moderate dysplasia with non-progression within 9 years (MODN9) and progression within 3 years (MODPC3)

In this cell classification case the most important features were the number of dark particles (DNO), followed by DAA and HUM3. The cell classification rate was 58.5 % for the progression group and 61.9 % for the non-progression group. 15/24 specimens of the progression group and 10/19 of the non-progression group were correctly classified with 10 unclear (Table 1). This result is also shown graphically in Fig. 1. In case without unclear decisions 19 and 14 cases, resp. were correctly classified. This result is significant for p < 0.001. Fig. 2 demonstrates 6 nuclei with high values (30 pixels, upper row), representing cells from patients with progression and low values (10 pixels, lower row) of DNO representing cells from patients with non-progression whereas the nuclear size and the mean optical density are similar.

To test whether MODPC9 belongs better to MODPC3 or to MODN9 all 7 specimens were classified as a testset according to the evaluated discriminant function. One of the cases was classified as MODPC3 the remaining as MODN9.

- Discrimination of patients with moderate dysplasia with progression within 5-9 years (MODPC9) and non progression within 9 years (MODN9)

By means of the subset of chromatin features, correct cell classification rates of 63.2 % for MODPC9 and 64.6 % for MODN9 were achieved. The best feature was NC13 which is the measure of correlation 2 of co-occurrence of flat texture, followed by RL2 and CO12. The subsequent specimen classification resulted in 6/7 correct MODPC9 decisions and 12/19 correct MODN9 decisions, 4 cases were falsely classified (Table 3). In case of no unclear decisions 4 cases remained falsly classified (p < 0.001).

- Three class discrimination of patients with moderate dysplasia with progression within 3 years (MODPC3), within 5-9 years (MODPC9) and non progression within 9 years (MODN9)

The number of dark particles was the most important feature to distinguish the three groups followed by DNOA and NC13. 14/24 MODPC3 specimens, all 7 of MODPC9 and 13 of MODN9 were correctly classified. 3 of the first progression group was classified into the non progression class MODN9 which is not favorable due to non adequate treatment planning. The cell and specimen classification results are listed in Table 4.

- Discrimination of patients with mild dysplasia with non-progression within 9 years (MIDN9) and progression within 3 years (MIDPC3)

NR1, CO6 and HUSPAN were the most important chromatin features to discriminate both classes. 71.3% of the nuclei of MIDPC3 and 69.1 % of MIDN9 were correctly classified. The subsequent specimen classification led to a classification rate of 87.5% (14/16) and 81.5% (22/27) respectively with two and five unclear decisions. Without the definition unclear cases 15/16 specimens of MIDPC3 and 24/27 specimens of MIDN9 were correctly classified (p<0.0001). The results are given in Table 5 and shown in Fig. 3.

- Discrimination of patients with mild dysplasia with non-progression within 9 years (MIDN9) and progression within 3-5 years (MIDPC5)

A cell classification rate for MIDPC5 of 68.3% and for MIDN9 of 60.4% were achieved using the chromatin features NC3, NR1 and CO6. This led to a specimen classification of 8/13 (61.5%) and 18/27 (66.7%) respectively with 3 false decisions and 11 unclear cases. The significance of this result without unclear cases was p<0.01 (Table 6).

- Discrimination of patients with mild dysplasia with non-progression within 9 years (MIDN9) and progression within 5-9 years (MIDPC9)

By means of the chromatin features MFRANG, MFM3 and CO14 a cell classification rate of 60.8% for MIDPC9 and 69.0% for MIDN9 could be achieved. The subsequent specimen classification rate without definition of unclear cases was 83.3% (10/12) and 77.8% (21/27) respectively. This result is significant for p<0.001. In Table 7 the specimen classifications are shown.

Summary of results:

These preliminary results suggest that using high resolution image analysis is a useful technique and highly sensitive for forecasting esophageal carcinoma in systematic patients. Our data shows that correct specimen classification rates of 72.5% to 90.7% could be achieved in case of mild and moderate dysplasia using chromatin features. These features can be applied to give hints for an individual treatment planning of each patient.

DISCUSSION

Esophageal cancer is a common malignancy with a very poor prognosis. The main reason for that is that most cases are asymptomatic until they are unresectable. Previous studies have shown that cancer chemoprevention has beneficial effects in reducing the progression of the esophageal carcinogenesis process and reduced the incidence rate of esophageal carcinoma around 50%. At present time, cancer chemoprevention is a rapidly expanding area of oncology. But it will be emphasized that clinical trials of chemoprevention require the unacceptable cost (millions of dollars), long duration (5-10 years), and large scale of effort (thousands of subjects) (7).

Even today, it is nearly impossible to get information on the deeper layers of the cells using the morphological analogy. This is partially a result of difficulties in discriminating the cells of different lesions and biological behavior, particularly in the discrimination of the nonspecific cell alternations and precancerous lesions from the invasive. In this setting, there is a clear need to develop practicable endpoint biomarkers in clinical trials of chemoprevention agents.

In the recent years, many publications on the subject of cell image analysis have appeared in the literature. But using high resolution image analysis to discriminate dysplasia with progression and non-progression is not well documented.

In 1966 Mendelsohn suggested that the recording of high resolution images and subsequent analysis by a digital computer could lead to the automated recognition and classification of five principal types of leukocytes (20). This classic paper applied new concepts to microscopic image analysis and statistical classifications. And then several authors initiated with the aim of objectivity to clinical cytology (30). In 1980 and 1981 Wied and Burger (3, 31, 32) have noted such subvisual clue as marker features for neoplastic events in the uterine cervix. Subtle changes in normal-appearing epithelial cells adjacent to malignant tumors are called malignancy associated changes (MAC) and can be detected by nuclear texture measurements. And then others found that marker features were of predictive value in assessing cases of moderate dysplasia in routinely prepared cervical smears (26). Such marker features have also been detected in other organ sites. Jahoda et al. (12, 13) correctly classified a very good result of normal and reactive mesothelial cells, histiocytes, and cells of metastatic adenocarcinoma of lung and breast origin in effusions and a high percentage of mesothelial cells and ovarian cancer cells in peritoneal fluids. Boon et al. (2) used image analysis to classify follicular adenoma and carcinoma of the thyroid on aspiration smears. Hutchinson et al. (11) achieved correct classification of prostate and breast aspirates by image analysis. Many researchers have used objective high resolution cytometry techniques to study precancerous lesions and malignant diseases from a wide variety of body sites. i.e. renal cell carcinoma (1), nevi and melanomas (8), lung (21) buccal mucosa (14), nasal mucosa (22) and head and neck (4). It is clear that image cytometry is capable of providing diagnostically and prognostically relevant information regarding disease. Particularly mentioned that F. Gao et al. (9) got surprisingly good results which used high resolution image analysis to discriminate esophageal severe dysplasia with progression and non-progression.

In recent years, some scientists emphasize and believe that use high resolution cytometric biomarkers of precancerous lesions which are urgently needed as endpoints in clinical trials of chemopreventive agents because they require less time, money and effort compared to the conventional endpoint of cancer incidence reduction (6, 7, 21).

In this study of 50 esophageal smears representing moderate dysplasia and 68 cases of mild dysplasia, it is shown that the progression to carcinoma within several years and non-progression can be discriminated using texture features (Table 2-7). Nearly 75% of the moderate dysplasia and about 80% mild dysplasia can be correctly classified by means of benign looking intermediate cells. We found that the behavior of both dysplasia types are similar. That is to say that some dysplasia lesions are not real precancerous lesions and are non-specific cell alternations. It is why that only a small proportion of dysplasia will progress to cancer, a majority of these lesions will spontaneously regress or stable. Our results have also shown that there has been little differences between three years and longer time of dysplasia with progression. We noticed similar classification rates for mild and moderate dysplasia discriminations. It is shown that even cells with subtle changes can be detected by high resolution image analysis.

To date, efforts at early detection of esophageal cancer and precancerous lesions have concentrated on cytologic or histologic categorization. During this period, nearly all dysplasia patients made diagnosis by cytologic and histologic criteria in clinical chemoprevention trial. But there was little possibility to get information on the deeper layers of the precancerous lesion cells by morphological methods. This has been proven to be difficult when depending on morphological observation only, since many cases have the same morphology but the biological behavior is quite different. Our results show that the texture features can forecast of precancerous lesions and can also be employed as surrogate endpoint biomarkers.

In prospective studies the preparation technique and staining conditions have to be standardized to avoid unexpected results. We believe that high resolution image analysis, which is a powerful weapon of cytology diagnosis and research can provide us with important information which is impossible to be observed through routine methods. Cell measurement of the texture features improves cytological diagnosis and could be used to monitor progression of lesions as well as the treatment.

REFERENCES

(1) Bibbo, M., Galera-Davidson, H., Dytoh, H.E., Gonzalez de Chaves, J., Lopez-Garrido, J., Bartels, P.H., Wied, G.L., Karyometry and histometry of renal-cell carcinoma, Anal. Quant. Cytol. Histol., 9: 182-187, 1987

(2) Boon, M.E., Trott, P.A., Kaam, H.V., Kurver, P.H.J., Baak, J.P.A., Computation of preoperative diagnosis probability for follicular adenoma and carcinoma of the thyroid on aspiration smears, Anal. Quant. Cytol. Histol., 4: 1-5, 1982

(3) Burger, G., Jütting, U., Rodenacker, K., Changes in benign cell populations in cases of cervical cancer and its precursors, Anal. Quant. Cytol., 3: 261-271, 1981

(4) Burger G, Aubele M, Clasen B, Jütting U, Gais P, Rodenacker K: Malignancy associated changes in squamous epithelium of the head and neck region. Analyt Cell Pathol 7:181-193, 1994

(5) Burger G, Jütting U: Specimen classification in cytometry: an intercomparison of various means of decision making. In: Gelsema ES, Kanal LN, eds. Pattern recognition in Practice II. Amsterdam: North-Holland Publ., 509-519, 1985

(6) Boone Ch, Kelloff G, Interaepithelial Neoplasia, Surrogate endpoint biomarkers, and cancer chemoprevention, J. Cellular Biochemistry, Supplement 17F: 37-48,1993

(7) Boone Ch, Kelloff G: Development of surrogate endpoint biomarkers for clinical trials of cancer chemopreventive agents: Relationships to fundamental properties of preinvasive (Intraepithelial) neoplasia, Journal of Cellular Biochemistry, Supplement 19: 10-22, 1994

(8) Fleming, M.G., Wied, G.L., Dytch, H.E., Image analysis cytometry of dysplastic nevi, J. Invest. Dermatol., 95: 287-291, 1990

(9) Gao F, Jütting U, Rodenacker K, Gais P, Lin P-Z, Relevance of chromatin features in the progression of esophageal epithelial severe dysplasia, Analyt. Cell. Pathol. 13: 17-28, 1997

(10) Kelloff G, Boone Ch, Steele V, Crowell JA, Lubet R, Sigman C, Progress in cancer chemoprovention: Perspectives on agent selection and short-term clinical intervention trials, Cancer Research (Suppl) 54: 2015s-2024s, 1994

(11) Hutchinson, M.L., Schultz, D.S., Stephenson, R.A., Wong, K.L., Harry, J., Zahnizer, D.J., Computerized microscopic analysis of prostate fine needle aspirates: Comparison with breast aspirates, Anal. Quant. Cytol. Histol., 11: 105-110, 1989

(12) Jahoda, E., Bartels, P.H., Bibbo, M., Bahr, G.F., Holzner, J.H., Wied, G.L., Computer discrimination of cells in serous effusions. I. Pleural Fluids, Acta Cytol., 17: 94-105, 1973

(13) Jahoda, E., Bartels, P.H., Bibbo, M., Bahr, G.F., Holzner, J.H., Computer discriminations of cells in serous effusions. II. Peritoneal Fluid, Acta Cytol., 17: 533-537, 1973

(14) Klawe, H., Rowinsk, J.: Malignancy Associated Change (MAC) in cells of buccal smears detected by objective image analysis, Acta Cytol. 18: 30-33, 1974

(15) Li, J.Y.: Epidemiology of esophageal cancer in China, Natl. Cancer Inst. Monogr., 62: 113-120, 1982

(16) Li, M.X., Li, B., Li, B.R., Recent progress in research on esophageal cancer in China, Adv. Cancer Res., 33: 173-249, 1980

(17) P.Z. Lin, Introduction of research works in esophageal precancerous lesions (dysplasia of esophageal epithelium), Chinese Journal of Cancer 5: 391, 1983

(18) P.Z. Lin, J.S. Zhang, Z.P. Rong, R. Han, S.P. Xu, R.Q. Gao, S.G. Cao, Second line prevention from esophageal cancer inhibitory therapy to block precancerous lesions, Chinese Journal of Cancer Research 1: 37-46, 1988

(19) P.Z. Lin, J.S. Zhang, Z.P. Rong, R. Han, S.P. Xu, R.Q. Gao, Z.W. Ding, J.X. Wang, H.J. Feng, S.G. Cao, Studies on medicamentous inhibitory therapy for esophageal precancerous lesions 3- and 5- year inhibitory effects of Antitumor-B, Retinamide and Riboflavin, Proc CAMS and PUMC, 5: 121-129, 1990

(20) Mendelsohn, M.L., Mayall, B.H., Prewitt, J.M.S., Bostrom R.S., Holcomb, W.G.: Digital transformation and computer analysis of microscopic image. In: Barer, R., and Cosslett, V.E. (eds.): Advances in Optical and Electron Microscopy. New York, Academic Press, 1968

(21) Palcic B.: Nuclear texture: Can it be used as a surrogate endpoint biomarker? J. Cellular Biochemistry, Supplement 19: 40-46, 1994

(22) Parkin, D.M., Pisani, P., Ferlay, J.: Estimates of the worldwide incidence of eighteen major cancer in 1985, Int. J. Cancer 54: 594-606, 1993

(23) Reith A.K, Reichborn-Kjennerud S., Aubele M, Jütting U, Gais P, Burger G, Biological monitoring of chemical exposure in nickel workers by image cytometry (ICM) of nasal smears, Analyt. Cell. Pathol. 6: 9-21, 1994

(24) Rodenacker K: Featuring of cellular objects. In: Burger G, Ploem JS, Goerttler K, eds. Clinical Cytometry and Histometry. Rep Part 3 4:73-78, 1973

(25) Rodenacker K: Invariance of textural features in image cytometry under variantion of size and pixel magnitude. Analyt Cell Pathol 8:117-133, 1995

(26) Rosenthal, D.L., Philippe, A., Hall, T.L., Harami, S., Missirlian, N., Suffin, S.C., Prognosis of moderate dysplasia. Predictive value of selected markers in routinely prepared cervical smears, Anal. Quant. Cytol. Histol., 9: 165-168, 1987

(27) Schmidt J, Aubele M, Jütting U, Rodenacker K, Luz A, Erfle V, Burger G: Computer-assisted imaging cytometry of nuclear chromatin reveals bone tumor virus infection and neoplastic transformation of adherent osteobast-like cells. Biochem Biophys Res Commun 164:728-735, 1989

(28) Shu, Y.J., The cytopathology of the esophagus; an overview of esophageal cytopathology in China, Acta Cytol., 27: 7-16, 1983

(29) Shu, Y.J., The Cytopathology of Esophageal Carcinoma, Masson, New York 1985

(30) Wied, G.L., Bahr, G.F., Oldfield, D.G., Bartels, P.H., Computer-assisted identification of cells from uterine adenocarcinoma. A clinical feasibility study with TICAS, Acta Cytol. 12: 357-370, 1968

(31) Wied, G.L., Bartels, P.H., Bibbo, M., Sychra., J., Cytomorphometric markers for uterine cancer in intermediate cells. Anal. Quant. Cytol., 2: 257-263, 1980

(32) Wied, G.L., Bartels, P.H., Dytch, H.E., Pishotta, F.T., Yamauchi, K., Bibbo, M., Diagnostic marker features in dysplastic cells from the uterine cervix, Acta Cytol., 26: 475-483, 1982

Table 1: Cytology results from the 1983 esophageal balloon cytology surveys in Linxian, China, the follow up results, and number of specimens and measured cell nuclei

NA: specimens not available

Table 2: Specimen classification results from moderate dysplasia with progression within 3 years (MODPC3) and non progression within 9 years (MODN9)

Specimen classification:

Specimen classification:

p < 0.001

Selected features with F-values:

DNO (83.5), DAA (83.5), HUM3 (27.7), HUCV (22.8), GCV ( 15.6)

Table 3: Specimen classification results from moderate dysplasia with progression within 7 to 9 years (MODPC9) and non progression within 9 years (MODN9)

Specimen classification:

Specimen classification:

p < 0.001

Selected features with F-values:

NC13 (78.4), RL2 (25.4), CO12 (31.9), HUCV (37.6), HLM3 ( 35.1)

Table 4: Cell and specimen classification results from moderate dysplasia in the 3 class case

Cell classification:

Specimen classification:

p < 0.001

Selected features with F-values:

DNO (67.1), DNOA ( 91.1), NC13 (26.8), GCV, (21.1), HAA (20.3)

Table 5: Specimen classification results from mild dysplasia cases with progression within 3 years (MIDPC3) and non progression within 9 years (MIDN9)

Specimen classification:

Specimen classification:

p < 0.0001

Selected features with F-values:

NR1, (492.3), CO6 (165.7), HUSPAN, ( 69.6), NC8 ( 80.9), FRCV (83.7)

Table 6: Specimen classification results from mild dysplasia cases with progression within 3-5 years (MIDPC5) and without progression within 9 years (MIDN9)

Specimen classification:

Specimen classification:

p < 0.01

Selected features with F-values:

NC3 (272.5), NR1 (63.2), CO6 (79.3), HCV (33.2), MM1 (20.1)

Table 7: Specimen classification results from mild dysplasia cases with progression within 5-9 years (MIDPC9) and non-progression within 9 years (MIDN9)

Specimen classification:

Specimen classification:

p < 0.001

Selected features with F-values:

MFRANG (259.6), MFM3 ( 48.2), CO14 ( 39.7), HUNO (63.7), DNOA (28.3)

Appendix:
List of features and their description:

CO6 Sum average of co-occurrence of extinction

CO12 Measure of correlation 1 of co-occurrence of extinction

CO14 Local mean of co-occurrence of extinction

DAA Relative area of dark particles

DNO Number of dark particles

DNOA Relative number of dark particles

FRCV CV of local factal dimension

GCV CV of gradient filtered image

HAA Relative area of bright particles

HCV CV of extinction of bright particles

HLM3 Skewness of lower skeleton

HUCV CV of upper skeleton

HUM3 Skewness of upper skeleton

HUNO Number of bright regions (particles)

HUSPAN Range of upper skeleton

MFRANG Range of local multifractal dimension

MFM3 Skewness of local multifractal dimension

MM1 Morphological parameter (first invariant moment)

NC3 Correlation of co-occurrence of flat texture

NC8 Sum entropy of co-occurrence of flat texture

NC13 Measure of correlation 2 of co-occurrence of flat texture

NR1 Short runs emphasis of runlength distribution of flat texture

RL2 Long runs emphasis of runlength distribution of flat texture

Figures