Chemoprevention with theaflavins of rat esophageal intraepithelial neoplasia quantitatively monitored by image tile analysis

The objective of the study was to compare three methods of monitoring the inhibition by dietary theaflavins of N-nitrosomethylbenzylamine (NMBA)-induced rat esophageal intraepithelial neoplasia: the mean tile grade, measured by computer-assisted quantitative image tile analysis; tumor multiplicity; and mean tumor size. A "tile" is defined as a small portion of a microscopic image at x 40, 87 x 292 microm in size. The computer divided the image of esophageal intraepithelial neoplasia into a grid of contiguous tiles and measured four tissue features within each tile based on cytonuclear and tissue architectural changes used by pathologists to diagnose intraepithelial neoplasia. The tile grade is defined as the weighted sum of the four feature measurements within a tile, the weights being determined by Fisher linear discriminant analysis. The mean tile grade of 300 tiles is used to grade rat esophageal intraepithelial neoplasia. NMBA was given s.c., 0.5 mg/kg, three times a week for 5 weeks. Theaflavins were given in the drinking water at 360 ppm (low dose) and 1200 ppm (high dose) throughout the experiment. In a given set of four groups of rats, one group received theaflavins alone, one NMBA alone, one NMBA plus low-dose theaflavins, and one NMBA plus high-dose theaflavins. One set of four groups, four rats/group, was sacrificed at the 15th week and another at the 20th week after starting NMBA; a final set with 15 rats/group was sacrificed at 25 weeks. At the 15th and 20th weeks, the mean tumor grade was the only variable that responded significantly (P < 0.01) to the low dose of dietary theaflavins. In fact, tumor multiplicity and mean tumor size sometimes showed enhancement at these doses. At the 25th week, when there were 15 instead of 4 rats/group, the mean tile grade, tumor multiplicity, and mean tumor size were all significantly (P < 0.01) decreased by both low and high doses of theaflavins. The mean tile grade is a more sensitive and reproducible variable than tumor multiplicity and mean tumor size in detecting the chemopreventive effects of theaflavins on intraepithelial neoplasia in the rat esophagus. This suggests that the mean tile grade may be a useful intermediate end point for use in human chemoprevention trials.

Quantitative grading of rat esophageal carcinogenesis using computer-assisted image tile analysis

Our objective was to grade, by computer-assisted quantitative image tile analysis, the intraepithelial neoplasia (also called dysplasia) that develops in esophagi of rats given N-nitrosomethybenzylamine (NMBA) for 5 weeks. To perform image tile analysis, the computer divides the video image of the neoplastic epithelium into a row of contiguous small rectangular images, or "tiles," 84 x 292 microm in size, and quantitatively measures four selected tissue features within each image tile. The computer then calculates a tile grade for each image tile as the weighted sum of the four feature measurements, transformed into statistical Z-scores, the weights being determined by Fisher linear discriminant analysis of 300 tile grades of the neoplastic epithelium referenced to the mean tile grade (MTG) of 300 image tiles of normal epithelium. The two grading parameters, MTG and the percentage of tile grades exceeding the MTG of normal epithelium by >4 SD units (%TG>4SD), were validated as endpoints for screening chemopreventive agents in the rat NMBA-induced esophageal carcinogenesis model in two ways: (a) after NMBA treatment, %TG>4SD developed in parallel with tumor incidence and tumor multiplicity (number of papillomas/tumor-bearing rat); and (b) placing the chemopreventive phenethylisothiocyanate in the food of NMBA-treated rats produced parallel reductions in MTG, tumor incidence, and tumor multiplicity. Both MTG and %TG>4SD, measured by quantitative image tile analysis, are sensitive and objective continuous parametric response variables expressed to three significant figures, with wide dynamic range, that may be evaluated by t tests to compare tissue neoplastic changes before and after treatment with a chemopreventive agent.

Image morphometric nuclear grading of intraepithelial neoplastic lesions with applications to cancer chemoprevention trials

A new image morphometric method of nuclear grading is described and assessed in the context of the evaluation of histological samples from ductal carcinoma in situ of the breast and cervical intraepithelial neoplasia. The method results in a continuous scaled variable, or nuclear grading scale, expressed in SD units from measured normal nuclei from breast or cervix. For a given histological preinvasive neoplastic lesion, the mean nuclear grade of measured nuclei was shown to be analogous to the histopathological nuclear grade of the same lesion assigned subjectively by the pathologist. In a chemoprevention trial of the effect of difluoromethylornithine given for 1 month to subjects with cervical intraepithelial neoplasia grade 3, pathologists could see no difference in 14 histological sections taken before and after difluoromethylornithine treatment. However, the image morphometric method detected a systematic effect of lowered mean nuclear grade and a decrease in the variability of nuclear grade expression. Twelve of 14 samples showed a lower posttreatment mean nuclear grade (P<0.05), and 13 of the 14 samples showed a decrease in the SD of their nuclear grade distributions (P<0.01). This study demonstrates the use of image morphometric nuclear grading in a chemoprevention setting. It may be very useful in supplementing the pathologist's histopathological grading by providing objective, quantitative assessments.

Quantitation of preinvasive neoplastic progression in animal models of chemical carcinogenesis

An assay method that precisely quantitates the cellular and tissue changes associated with early, preinvasive neoplasia is much needed as a surrogate endpoint biomarker (SEB) in clinical trials to predict the potential efficacy of chemopreventive agents in bringing about cancer incidence reduction. Quantification of histological changes at the tissue level are potentially powerful SEB's since these visually apparent changes are common in all neoplastic development, regardless of tissue type or neoplastic cause. Currently, subjective inspection of the histological appearance of sectioned and stained material, or "grading," by experienced pathologists is used to evaluate neoplastic progression. This has well-known limitations of reproducibility, accuracy, and resolution of grading scale. Since neoplastic changes are visually apparent and morphologic in nature, quantification by image analysis is a measurement modality of choice. Image analysis was implemented through the use of high-resolution "tiled" images of complete tissue sections. A histological grading system, or "scale," was developed that could be expressed in terms of normal deviate units of multiple and different morphometric descriptors. Neoplastic growth was characterized quantitatively with multiple measurements on each tissue image tile, which were combined into a single number for each tile, i.e., a histologic grade per tile, and parameters from the distributions of these measurements were used to represent the histologic grade for the entire region considered. This concept provided a uniform final scale in similar units of measurement, regardless of which tissues were graded. Also, the grading scale automatically adjusted measurement variance for different tissues by using normal tissue for each different type to obtain the normalization to standard deviation (z) units. This further defined a uniform final scale and maintained standard references. Using this method, results from two well-known animal models of carcinogenesis, squamous cell carcinoma of SENCAR mouse skin induced by benzo(a)pyrene (B[a]P), and squamous cell carcinoma of the rat esophagus induced by N-nitrosomethylbenzylamine (NMBA), were compared to each other. Image analysis was performed on skin tissue sections from a total of 64 SENCAR mice, and esophagus tissue sections from 96 Fischer-344 rats. In both cases, a quantitative expression of the preinvasive neoplastic response to the carcinogen as a function of time of exposure was expressed along a continuous grading scale in standard deviation units (z). In the SENCAR mouse skin animal model, similar cohorts of 4 mice at 20 weeks showed significant modulation of B[a]P-induced neoplasia by treatment with the antiproliferative agent difluoromethylornithine, P < .05. In the rat esophagus animal model, similar cohorts of 6 rats at 10 and 15 weeks showed significant modulation of NMBA-induced neoplasia by treatment with the antimutagen phenethyl isothiocyanate, P < .05.

Properties of intraepithelial neoplasia relevant to the development of cancer chemopreventive agents

Cancer chemoprevention is concerned with the development of drugs or diet supplements that will avert the onset or stop the progression of the intraepithelial neoplasia which precedes invasive cancer. Two basic processes underlie the onset and development of intraepithelial neoplasia. First is genomic instability (often associated with chronic diffuse epithelial hyperplasia), which is the increased production of genomic structural variants due to unrepaired DNA breaks with secondary formation of abnormal structures, including "mutator" mutations in genes responsible for genomic stability, gene copy amplification or loss from DNA breakage-fusion-anaphase bridge cycles, unequal sister chromatid exchange, and accumulation of double minutes. Second is the development within an epithelium having genomic instability of multicentric neoplastic lesions that independently progress through each of the following processes at a continuously accelerating rate: clonal evolution, hyperproliferation, production of genomic structural variants, and apoptosis. Recommended chemoprevention strategies based on these mechanisms are (1) early diagnosis and treatment of genomic instability before the appearance of intraepithelial neoplasia, i.e., during the "predysplastic" or "premorphologic" phase, (2) development of multiple agents that block intralesional proliferation at steps along the "command" pathways of mitotic signal transduction and along the "execute" pathways of synthesis of daughter cell components, (3) development of nontoxic antiinflammatory agents, antioxidants, antimutagens, and proapoptotics, (4) avoidance of "clonal escape" through use of drug combinations, and (5) use of computer-assisted quantitative image analysis to assay modulation of surrogate endpoints in chemoprevention clinical trials.

Properties of intraepithelial neoplasia relevant to cancer chemoprevention and to the development of surrogate end points for clinical trials

Cancer chemoprevention is defined as the prevention of cancer by the administration of diet supplements or drugs. A drug discovery effort should therefore focus on finding agents that will avert the process of intraepithelial neoplasia which precedes invasive cancer. Over 30 agents developed by the chemoprevention program at the National Cancer Institute are being tested against intraepithelial neoplasia of many organ sites in more than 80 clinical trials. Two basic mechanisms underlie the onset and development of intraepithelial neoplasia. First is the development of the two precursor lesions of chronic diffuse epithelial hyperplasia and genomic instability, the latter being produced by "mutator" mutations in genes responsible for genomic stability, by gene copy amplification or loss from DNA breakage-fusion-anaphase-bridge cycles, by unequal sister chromatid exchange, and by accumulation of double minutes. Second is the development of multicentric intraepithelial neoplastic lesions which independently progress through each of the following processes at a continuously accelerating rate: clonal evolution, hyperproliferation, production of genomic structural variants, and apoptosis. Recommended chemoprevention strategies based on these mechanisms are (i) the development of better technology for early diagnosis, (ii) the development of multiple agents that block intralesional proliferation at steps along the signal pathway of mitotic signal transduction and along the signal pathway of synthesis of daughter cell components, (iii) the development of nontoxic anti-inflammatory agents, anitoxidants, antimutagens, and proapoptotics, (iv) the avoidance of "clonal escape" through use of drug combinations, and (v) the use of computer-assisted quantitative image analysis to assay modulation of surrogate end points in chemoprevention clinical trials.

Cytology, flow cytometry, image analysis, and interphase cytogenetics by fluorescence in situ hybridization in the diagnosis of transitional cell carcinoma in bladder washes: a comparative study

The diagnosis of transitional cell carcinoma (TCC) in bladder washes is a diagnostic challenge to cytology. This study assessed the role of flow cytometry (FCM), image analysis (IA), and interphase cytogenetics by fluorescence in situ hybridization (FISH) as adjuncts in the cytodiagnosis of TCC in bladder washes. Forty separate samples of bladder washes were prospectively evaluated by conventional cytology (CY), FCM, IA, and FISH, and the results were compared with the subsequent surgical biopsy specimens which revealed 26 TCC (3 GR I, 6 GR II, 17 GR III) and 14 benign lesions. Using histology as the "gold standard" and following the previously published criteria for detection of TCC by CY, FCM, IA, and FISH, the concordance rates between histology and CY, FCM, IA, and FISH were 75, 74, 89, and 83%, respectively. CY, FCM, IA, FISH, and histology were concordant in 54% of the cases. The sensitivity of CY, FCM, IA, and FISH were 61, 72, 91, and 73%, respectively, while the specificity were 100, 80, 83, and 100%, respectively. The combined sensitivity of all the parameters was 96%. Interestingly, the false positive cases by FCM and IA showed cystitis. We conclude that IA has the highest sensitivity in detecting TCC in bladder washes followed by FISH, FCM, and CY, while CY and FISH have the highest specificity. This study indicates that FCM, IA, and FISH are useful adjuncts to cytology in the diagnosis of TCC in bladder washes. The finding of DNA-aneuploidy in cystitis warrants further investigation.

Cervical cell recognition and morphometric grading by image analysis

Cervical cell recognition by morphometric image analysis was compared to human visual cell recognition on the same 6,375 cells from 40 dysplastic, CIS, invasive, and 10 normal pap smears. The experimental approach defined receiver operating characteristic (ROC) curves for morphometric image analysis which could be rigorously compared to previously established human visual cell recognition ROCs on the same cells. Overall performance was measured Az, the area under the ROC curves in the two instances. For morphometric image analysis cell recognition, Az = 0.91, and for human visual cell recognition, Az = 0.87. These results clearly demonstrated that morphometric image analysis is equivalent to experienced human observers in ability to recognize isolated cells from cervical smears. An approach was also developed to link the ROC analytic methods of this study to a cytopathological or histopathological grading system, or "scale", that could be expressed in terms of normal deviate units of morphometric descriptors. This approach has the advantage of describing the grading scale in terms of its ROC characteristics; in essence, it describes performance for that grading scale at any decision point along the scale, if used for two-category classification. Additionally, This concept provides for a uniform final scale, regardless of which cells or tissues are graded. Also, this type of grading scale would automatically adjust itself for measurement variance for different types of cells or tissue, by reference to normal cells or tissues, so that a standard reference could be maintained.

A method of correcting DNA ploidy measurements in tissue sections

A new method of performing DNA ploidy measurements in tissue sections using image cytometry is described. The method involves an image processing "object" filtering operation to remove cut and overlapping nuclei, and DNA correction of larger nuclei for the part of the DNA cut away. The methodology of the technique is developed in detail, and the results of testing using sections of rat liver are presented. The results indicated reliable correction of DNA ploidy histograms to reconstruct the polyploid nature of this material. The sensitivity of the material to sectioning thickness errors and methods to overcome this were shown, along with an example of its use on prostate bioptic gun tissue sections.

Quality control in image cytometry: DNA ploidy

DNA ploidy has become a commonly performed quantitative image cytometry test in microscopic pathology. This has led to the need to develop quality control procedures to aid in assuring uniform and reliable test results. There are a number of unique issues related to the emerging technology of image analysis and its routine use as a quantitative microscopic assay that require consideration before establishing a quality control program. Previous considerations of this topic have primarily related to measurement issues, e.g., accuracy comparisons to other methodologies, calibration of instrumentation, sources of measurement error, and the interpretation of measurements results. Although these issues are critically important, with more routine usage the focus is now turning to quality control of the overall testing process in everyday use. Control charts and methods of controlling the total measurement process such as have been used in clinical chemistry, need to be established. As one of the first image assays in pathology, quality control procedures established now for DNA ploidy measurements could help shape the development of this field, especially as pathology transitions from being a subjective visual microscopic inspection process to a quantitative measurement process. This paper discusses these issues as they relate to overall quality assurance for the DNA ploidy test and describes a quality control program developed for an active breast cancer testing laboratory specializing in image cytometry tests, including DNA ploidy. The quality control program includes calibration control charts, control charts for internal diploid controls, check samples, and computerized individual histogram interpretation.

Chromatin texture measurement by Markovian analysis. Use of nuclear models to define and select texture features

The use of nuclear grade as a prognostic indicator in breast cancer has been limited by its poor interobserver reproducibility. Automated cell classification using digital image analysis is one approach to this problem. Nuclear chromatin distribution, an important feature used in nuclear grading, can be quantitated with texture analysis. Markovian analysis is one method of analyzing texture features that is available in a commercially available image analysis system, the CAS-100. In order to select optimal Markovian features for use in nuclear grading of breast cancer, 16 nuclear models were created with computer graphics that demonstrated specific components of nuclear chromatin pattern, such as granularity, contrast, symmetry, peripheral chromatin clumping, and number and shape of nucleoli. These models were analyzed on the CAS-100 image analysis system using software capable of measuring 22 Markovian texture features at 20 levels of pixel resolution (grain). We were able to show that Markovian analysis performed well in discriminating between degrees of chromatin granularity (finely vs. coarsely clumped), amount of contrast (vesicular change), thickness of peripheral chromatin and number of nucleoli. Of the 22 Markovian features, 10 were selected as optimal for discriminating between the above chromatin patterns. Similar optimal Markovian features were found when measurements were performed on captured images of breast cancer cells. The use of these selected Markovian texture features may allow a more rational approach to the use of image analysis for cell classification.

HER-2/neu oncogene expression and proliferation in breast cancers

Amplification of the HER-2/neu proto-oncogene in breast cancer has been reported to correlate with poor patient prognosis. The proliferation, or growth fraction, of cells has also been shown to be of prognostic importance in breast cancer. A study was conducted to evaluate the correlation between HER-2/neu gene expression and proliferation in breast cancer. Quantitative immunohistochemical methods for the detection of the HER-2/neu protein expression and for assessing the proliferation fraction on frozen sections of tumor cells were used. The detection of epidermal growth factor receptor (EGFR) along with quantitative DNA ploidy analysis, also was performed on the same breast cancers. The results indicated two subgroups of invasive ductal carcinoma; 1) HER-2/neu overexpressing cases that were negative for EGFR expression and had low proliferation fraction, and a tetraploid DNA pattern (22 cases), and 2) other combinations of HER-2/neu expression and EGFR expression, with a high proliferation fraction and an aneuploid DNA pattern (38 cases). Eight cases of carcinoma in situ were positive for HER-2/neu overexpression and negative for EGFR expression, and had a high proliferation fraction and a tetraploid DNA pattern. Twenty-six cases of low-grade carcinoma exhibited low proliferation and a diploid DNA pattern.

HER-2/neu oncogene expression and DNA ploidy analysis in breast cancer

This study was performed to evaluate the correlation between HER-2/neu gene expression and DNA ploidy patterns. Forty-five cases of breast-cancer were analyzed. Immunohistochemical staining of HER-2/neu protein on frozen sections was used to detect the HER-2/neu protein, and the Feulgen DNA staining method was used to assess DNA amounts in the same tumor cells. Positive HER-2/neu overexpression was evaluated visually, and quantitation of the HER-2/neu protein was measured by image analysis. Twenty-two of the 45 cases were visually scored to be positive for the overexpression of the HER-2/neu protein, and these cases also contained above 10% HER-2/neu protein compared with a standard control cell line. All 22 of these cases had near-tetraploid DNA content. In contrast, cells, derived from the 23 cases that did not overexpress the HER-2/neu protein, contained DNA amounts that ranged from euploid (diploid) to varying degrees of aneuploid. The results of this study indicated that tumors that overexpress the HER-2/neu protein have tetraploid or near-tetraploid DNA content. This pattern could relate to the biological behavior of these tumors.

Comparison of the Papanicolaou and Feulgen staining methods for DNA quantification by image analysis

The possibility of using archival cytology material to study the evolution of neoplastic disease with regard to DNA content abnormalities was investigated. The accuracy of measuring the integrity optical density (OD) of nuclei that correlates to DNA amounts of those nuclei, on slides stained by the Papanicolaou method, was assessed and compared with a standard Feulgen method. Our data on rat liver nuclei peritoneal washings from patients with ovarian cystadenofibromas and ovarian cystadenocarcinomas suggested that analysis of cytological material using the Papanicolaou method is not reliable and that destaining the slides followed by Feulgen staining provides an optimal and reliable method of DNA quantification.

Biological grading of breast cancer using antibodies to proliferating cells and other markers

Quantitation of immunohistochemical staining by image analysis was performed on 50 breast cancers stained with the monoclonal antibody Ki-67 to determine the growth fraction and its correlation with tumor grade. A high degree of correlation was shown. For each case the DNA ploidy was determined by quantitation of the DNA Feulgen stain by computerized microdensitometry. DNA content of breast tumor cells correlated to the histopathologic grade at which poorly differentiated tumors are more likely to be aneuploid. Quantitation of immunohistochemistry for estrogen and progesterone receptors had a high degree of correlation with the steroid binding assay, such as dextran-coated charcoal assay (DCCA), and were weakly correlated to histologic grade. In summary, our results indicated that quantitation of Ki-67-positive nuclear area and of DNA content by image analysis provides an objective method for assessing tumor cell growth fraction and DNA ploidy. Quantitation of steroid receptors by immunohistochemistry is a better and easier technique than those currently used to determine the best therapy for postmenopausal women. These methods can be performed on small frozen sections or needle aspirates in quantities that are insufficient for current steroid binding assays. Thus, this method is prognosticly useful even for patients with small breast lesions.

The evaluation of estrogen receptor in primary breast carcinoma by computer-assisted image analysis

A monoclonal antibody prepared against estrogen receptor has been shown to be specific and sensitive for the detection of estrogen receptor in human breast lesions by use of immunohistochemical methods. Two hundred selected cases of primary breast carcinoma were assayed for estrogen receptor content by biochemical and immunohistochemical procedures. Quantitative evaluation was by biochemical, immunohistochemical, and automated computer-assisted image analysis using the Cell Analysis System's CAS/100 machine (Lombard, IL). Quantitative estrogen receptor content was determined by dextran-coated charcoal analysis and sucrose density gradient analysis. Immunohistochemical evaluation incorporated both intensity and distribution of staining, yielding a subjective score, histologic score (HSCORE). An objective quantitation, also incorporating intensity and distribution of staining, was done by computer-assisted image analysis, quantitative immunocytochemical score (QIC SCORE). HSCORE analysis was done with and without methyl green counterstain with no loss of sensitivity. Comparison of QIC SCORE with the biochemical and immunohistochemical analysis of the tissues examined revealed excellent sensitivities and specificities. These data suggest that automated image analysis provides an effective qualitative and quantitative means of evaluating estrogen receptor content in human breast cancers.

Optical microscope system for standardized cell measurements and analyses

A fully integrated optical microscope and computer workstation for the pathology laboratory is described along with a system module for that workstation, the Cell Measurement Program (CMP). This module allows for the acquisition and storage of digitized microscope images; measurement of a standard set of cell features, or descriptors, calibrated for accurate densitometry; and a comprehensive set of statistical analyses and display procedures. This system is useful in research in cell biology and in cancer research, allowing the investigator to use the microscope as a measuring instrument.

Malignant cell detection and cervical cancer screening

A study was conducted to determine the isolated, single-cell detection characteristics of human observers as this information relates to cervical cancer screening. Two interrelated experiments were performed. First, the receiver operating characteristic (ROC) was obtained for slide screening. In this experiment, approximately 1,200 slides were examined. Second, ROCs were obtained for human observer cell discrimination, using a rating method. An individual's curves were computed, assuming a multiple decision criterion. In this experiment, 6,375 cells from the same specimens used in the slide screening experiment were studied. In both experiments, results were analyzed using a Gaussian signal-detection model. This approach provided analytical detection criteria and rigorous definition of ROCs. These experiments addressed the problem of where the screening information lies: (1) in individual cells alone or (2) with additional components in global or other a priori information. We quantified the detection requirements of (1) the Papanicolaou smear screening process, Az = 0.99, and (2) the capabilities of trained cytotechnologists on isolated single cells, Az = 0.87. System modeling using intermediate cell detection instead of "rare event" detection resulted in a reduction of the predicted number of cells required for analysis from approximately 60,000 to 750.

The brain as a sponge: a computed tomographic look at Hakim's hypothesis

Hakim's hypothesis, which explains the brain's adaptation to pressure in hydrocephalus and the enlargement of ventricular volume, was studied with computed tomographic (CT) scanning. The critical element in the theory is that brain water decreases when there is a pressure gradient within brain tissue. Extraaxial hematomas uniformly increase the CT density of adjacent brain tissue, and this is best explained by water shifts out of the tissue. In hydrocephalus cases studied early after shunting (10 days or less), a decrease in ventricular volume is accompanied by an overall decrease in the CT density of the brain. These CT changes support Hakim's view of the brain as a sponge that can change its hydration under pathological conditions.

Scene-segmentation algorithm development using error measures

Development of scene-segmentation algorithms has generally been an ad hoc process. This paper presents a systematic technique for developing these algorithms using error-measure minimization. If scene segmentation is regarded as a problem of pixel classification whereby each pixel of a scene is assigned to a particular object class, development of a scene-segmentation algorithm becomes primarily a process of feature selection. In this study, four methods of feature selection were used to develop segmentation techniques for cervical cytology images: (1) random selection, (2) manual selection (best features in the subjective judgment of the investigator), (3) eigenvector selection (ranking features according to the largest contribution to each eigenvector of the feature covariance matrix) and (4) selection using the scene-segmentation error measure A2. Four features were selected by each method from a universe of 35 features consisting of gray level, color, texture and special pixel neighborhood features in 40 cervical cytology images . Evaluation of the results was done with a composite of the scene-segmentation error measure A2, which depends on the percentage of scenes with measurable error, the agreement of pixel class proportions, the agreement of number of objects for each pixel class and the distance of each misclassified pixel to the nearest pixel of the misclassified class. Results indicate that random and eigenvector feature selection were the poorest methods, manual feature selection somewhat better and error-measure feature selection best. The error-measure feature selection method provides a useful, systematic method of developing and evaluating scene-segmentation algorithms.

Red blood cell morphology in sickle cell anemia as determined by image processing analysis: the relationship to painful crises

Red blood cell morphology was studied in the peripheral blood of adults with sickle cell anemia to determine if changes occur during painful crises. Image processing of the cells with an automated system of red blood cell analysis was used. Four groups of cells were observed: normocytes, macrocytes, target cells, and cells with the shape of irreversibly sickled cells. During asymptomatic periods, the percentages of these cells differed in each individual but were typical for that individual and generally were stable. During crises, macrocytosis occurred and the concentration of irreversibly sickled cells showed greater fluctuation. The macrocytosis most likely reflected a marrow response to increased hemolysis and demonstrated that the increased red blood cell destruction observed during pain crises may be more extensive than previously considered. Changes in the concentration of irreversibly sickled cells during crises were not consistent and could not be used as an indicator of a crisis. Image processing with automated red blood cell analysis allows for accurate assessment of all the morphologic groups of red blood cells in patients with sickle cell anemia and compares well with standard methods for measuring the concentration of irreversibly sickled cells.

Microspectrophotometric studies of Romanowsky stained blood cells. II. Comparison of the performance of two standardized stains

This paper describes a comparison of the performance of two standardized Romanowsky blood stains, namely those of Marshall et al. and Wittekind et al., both containing azure B and eosin alone. Stain performance is assessed objectively by the use of three complementary techniques, all based on the visible absorbance spectra of stained cellular substrates. The first of these techniques is a simple comparison of the shapes and heights of the absorbance spectra. The second technique uses the CIE Colorimetric System, and thus permits the quantitation of colour in a manner that agrees with human observation. CIE co-ordinates (chromaticity points, luminance) are calculated directly from absorbance spectra. The third technique is that of spectral subtraction, which yields a set of factors which describe the quantities of component dyes which are bound by the object. This technique, unlike the other two, requires a priori knowledge of the dyes used in the stains, and their spectra when bound to cellular substrates. Although the differences between the two methods are subtle, and hard for the subjective observer to define, the objective methods described here do show statistically significant differences. Wittekind's stain produces less intense staining, except for lymphocyte and monocyte cytoplasms. To the human eye, the differential coloration of these two substrates is more pronounced, but the difference between all nuclei and cytoplasm is less marked. The major difference in the uptake of dye components is in the small quantities of eosin dimer that are bound in this technique.

Assessment of the anemia of chronic disease by digital image processing of erythrocytes

Eighteen patients with the anemia of chronic disease were studied to determine the usefulness and accuracy of image processing analysis of erythrocytes for diagnosis. Diagnostic tests used for comparison included estimates of serum ferritin, serum iron and iron-binding capacity, stainable marrow iron, and erythrocyte morphology as determined by standard methods. The findings show that the analyses obtained by image processing were diagnostic for 83% of the patients with chronic disease. Serum ferritin levels were supportive of diagnosis for 33% of the patients, and serum iron levels were useful for approximately 25% of the patients. Cell indices and marrow iron were of limited value. The study demonstrates that quantitative information obtained by digital image processing of erythrocytes can be very useful for the diagnosis of the anemia of chronic disease.

Microspectrophotometric studies of Romanowsky stained blood cells. I. Subtraction analysis of a standardized procedure

This paper describes a microspectrophotometric study of blood smears stained by a simple, standardized Romanowsky technique, using only the dyes azure B and cosin. Absorbance spectra are presented for twenty-two classes of cellular object, and for the two dyes in solution, together with tabulations of spectral maxima, and suitable wavelengths for use in automated image processing. The colours of objects stained with azure B/eosin are discussed in terms of absorbance spectra. By a spectral subtraction technique, it is shown that the differential colouration of various cell structures may be explained satisfactorily in terms of the varying proportions of only four dye components. These are the monomers and dimers of azure B and eosin. Polymerization was found to occur both in solution and on binding to biopolymers. A similar analysis of a conventional Romanowsky stain would present much greater difficulties, due to the greater number of dye components, which, however, contribute little to the colours observed.

Quantitative measurement of red blood cell central pallor and hypochromasia

A quantitataive definition and techniques of measurement for central pallor of red blood cells are proposed. These are based on high-resolution measurements of absorbance across the center of the cell. Thus, the measurements reflect both variations in cell thickness and hemoglobin concentration. Although contributions of thickness and concentration may differ in individual cells, to a first approximation, a specific cell may be considered as having a similar concentration of hemoglobin throughout, and thus the major contribution to the central pallor is that due to the difference in thickness between the edges of the cell and the center. The definition proposed expresses central pallor as the percentage volume of indentation, comparing the red cell to a disc of uniform absorbance equal to the maximum found at the cell edges. Population distributions of central pallor then provide a basis for quantitation of hypochromasia. The mean and standard deviation of such distributions are proposed as quantitative descriptors. Sample distributions from 27 normal persons, 8 patients with spherocytic anemia and 26 patients with iron deficiency anemia were studied.

Studies on Papanicolaou staining. I. Visible-light spectra of stained cervical cells

This paper presents visible-light absorbance spectra of the nuclei and cytoplasms of ten types of cervical cells stained by the Papanicolaou technique, together with tables summarizing the data. It is concluded that for automated image processing of Papanicolaou smears the most suitable wavelengths are 531 nm to maximize the contrast of cell against background and a wavelength in the range 560 nm to 605 nm (with a mean at 575 nm) to maximize the contrast of the nucleus against the cytoplasm.

Studies on Papanicolaou staining. II. Quantitation of dye components bound to cervical cells

Dye binding to Papanicolaou-stained cervical cells was quantitated with a spectral subtraction technique. It was shown that of the dye components of Papanicolaou stains only aluminum hematein, orange G, light green SF and eosin Y are bound to cervical cells and that their chromophores do not interact. Bismarck brown Y was not bound to cervical cells, which confirms the subjective findings of previous authors, and so it should be omitted. No evidence of substrate-related variations in dye polymerization was found. The proportions of the four dyes were found to vary considerably from substrate to substrate. Spectra are presented of (1) solutions of the component dyes of the Papanicolaou stain, and (2) cervical cells stained with these individual component dyes, for comparison with the full Papanicolaou technique.

Error measures for objective assessment of scene segmentation algorithms

Scene segmentation is an important element in pattern recognition problems. Previous efforts to evaluate and compare scene segmentation procedures have been largely subjective. Quantitative error measures would facilitate objective comparison of scene segmentation algorithms. A theoretical discussion leading to a new generalized quantitative error measure, G2, based on comparison of both pixel class proportions and spatial distributions of "true" and test segmentations, is presented. This error measure was tested on 14 manual segmentations and 40 gynecologic cytology specimens segmented with five different scene segmentation techniques. Results indicate that G2 seems to have the desirable properties of correlation with human observation, categorization of error allowing for weighting, invariance with picture size and ease of computation necessary for a useful scene segmentation error measure.

Automated scene analysis of CT scans

Since the advent of computed tomography, there has been an increasing realization that CT scans contain quantitative as well as qualitative information useful in the diagnostic process. Often however, the use of this information is impeded by the tedious manual outlining of the areas of interest in the scan. To alleviate this problem, we have developed a scene segmentation algorithm which will automatically delineate areas of interest in a CT scan. This procedure uses known information about the expected objects in the scan in conjuction with an algorithm to label those objects. The resultant scene segmentation divides the scan into four anatomical areas: skull, normal brain, high density lesions and CSF. After an area of interest is interactively selected by the clinician, volume, density or other quantitative measures may be computed. Limitations of the algorithm and its clinical applications are discussed.

A multi-spectral approach for scene analysis of cervical cytology smears

A multi-spectral approach for the scene analysis of cervical cytology smears, using multiple images of a scene digitized through suitably chosen color filters matched to the Papanicolaou stain, has been proposed here. This technique involves clustering of two-dimensional data for extracting cytoplasm of the epithelial cells. Its performance on an experimental data set of 233 scenes involving more than 10 types of normal and malignant epithelial cells has been compared with density and gradient thresholding techniques. This resulted in an approximate 83% rate of success compared to approximately 40% for the rest of the other techniques.

An automated method of differential red blood cell classification with application to the diagnosis of anemia

A method of automated red cell analysis suitable for the rapid classification of large numbers of red cells from individual blood specimens has been developed, and preliminarily tested on normal bloods and clinically proven cases of anemias and red cell disorders. According to this method digital image processing techniques provide several features relating to shape and internal central pallor configurations of red cells. These features are used with a fully automated decision logic to rapidly provide a quantitative "red cell differential" analysis, a report of the percentage subpopulations of recognized categories of red cells. For each subpopulation, measurements of mean cell area, mean cell hemoglobin content and mean cell hemoglobin density are provided. The nine types of red cell disorders studied with this method were: (a) iron deficiency anemia, (b) the anemia of chronic disease, (c) beta-thalassemia trait, (d) sickle cell anemia, (e) hemoglobin C disease, (f) intravascular hemolysis, (g) hereditary elliptocytosis, (h) hereditary spherocytosis, and (i) megaloblastic anemia due to folic acid deficiency. Preliminary indications are that the red cell differential is useful in distinguishing between these conditions.

Automated classification of blood cell neutrophils

The classification of white blood cell neutrophils into band neutrophils (bands) and segmented neutrophils (segs) is a subproblem of the white blood cell differential count. This classification problem is not well defined for at least two reasons: (a) there are no unique quantitative definitions for bands and segs and (b) existing definitions use the shape of the nucleus as the only discriminating criterion. When cells are classified on a slide, decisions are made from the two-dimensional views of these three-dimensional cells. A problem arises because the exact shape of the nucleus becomes indeterminate when the nucleus overlaps so that the filament is hidden. To assess the importance of this problem, this paper quantitates the classification errors due to overlapped nuclei (ON). The results indicate that, using only neutrophils without ON, the classification accuracy is 89%. For neutrophils with ON, the classification accuracy is 65%. This suggests a classification strategy of first classifying neutrophils into three categories: (a) bands without ON, (b) segs without ON and (c) neutrophils with ON. Category III can then be further classified into segs and bands by other stretegies.

Image processing for automated erythrocyte classification

Digital image processing and pattern recognition techniques were applied to determine the feasibility of a natural n-space subgrouping of normal and abnormal peripheral blood erythrocytes into well separated categories. The data consisted of 325 digitized red cells from 11 different cell classes. The analysis resulted in five features: (a) size, (b) roundness, (c) spicularity, (d) eccentricity and (e) central gray level distribution. These features separated the data into six distinct condensed subgroups of red cells. Each subgroup consisted of morphologically similar cells: (a) macrocytes, (b) normocytes, (c) schistocytes, acanthocytes and burr cells, (d) microcytes and spherocytes, (e) elliptocytes, sickle cells and pencil forms and (f) target cells. The concept of a quantitative "red cell differential" was introduced, utilizing these subgroup definitions to establish subpopulations of red cells, with quantifiable indices for the diagnosis of anemia, at the specimen level.

Erythrocyte morphology and centrifugal "spinner" blood film preparations

The centrifugal "spinner" blood film preparation technique was investigated with respect to the preparation of blood films with the retention of conventional erythrocyte morphology. It was determined that the initial spinning produced a monolayer blood film while the duration of spinning affected the separation of the red cells in the monolayer and altered their individual morphologic characteristics. The spinner preparation resulted in a slide with the total surface similar in morphologic characteristics to one thin slice, or cross-sectional area, across a typical wedge preparation. The spinning time in effect determined the position of the cross-section. In many series of slides from the same blood, spun at increasing time intervals, at least one slide in each series contained erythrocytes whose morphology matched the "best" area of the corresponding wedge slide. The effects of red cell concentration and plasma viscosity were varied experimentally, to assess independently the effect of each on monolayer formation and morphology. The results indicated that it was necessary to spin for a longer time to overcome the viscosity effects due to the red cell concentration and also those due to plasma. Two evaluation criteria, "separation" and "central pallor development," were shown to produce different spinning time estimates. These differences were more pronounced at higher blood viscosities than at lower blood viscosities.