Search

The malignant melanoma (MM) is the most dangerous human skin disease. The incidence increased dramatically during the last years. The only chance for the patient is an early recognition and excision of the MM. The best diagnostic method for this is skin surface microscopy or dermatoscopy. Its use, however, requires much expertise. In order to support learning and using the method, a computer-based dermatoscopy workstation is being developed. Among others, new complexity measures are used for the image analysis.

Background: Photographic documentation of melanocytic skin lesions is important. Storage and retrieval of slides, however, take much time and space., Objective: Our purpose was to develop and clinically test a computerized acquisition and surveillance (CAS) unit with a television camera for monitoring including measurements of lesional areas., Methods: A CAS unit connected with a skin surface microscopic television camera was used for monitoring of melanocytic nevi (MN). The lesional area and the skin surface microscopic appearance (SMA) were analyzed after 10 to 21 months in 54 of 1355 MN., Results: In 19 MN (35.2%), changes were found. In eight cases, changes in size of more or less than 15% were detected; in five cases only the SMA changed. In six cases both characteristics changed., Conclusion: In approximately 25% of MN, changes were only detectable in the SMA but not with area measurements. This favors the use of systems such as CAS because only they allow a time-saving comparison of actual and previous images.

To elucidate the reasons for the malignant histologic appearance of melanocytic nuclei within benign Spitz nevi (SN), we evaluated nuclear DNA distribution and nuclear size using a computerized image analysis system. In each case of 28 SN and 34 malignant melanomas (MM), about 100 randomly sampled nuclei were analyzed, prepared as monolayers from paraffin-embedded tissues. Large nuclei in MM (nuclear area > mean nuclear area of normal melanocytes + 4 delta) were significantly more likely to be aneuploid (DNA content > or = 5c) than large nuclei in SN chi2 test, p < 0.0001). Only two of 990 large SN nuclei exhibited DNA values higher than 5c, whereas 236 of 2,024 large MM nuclei were aneuploid or polyploid. Accordingly, in multivariate analysis, five features of DNA distribution proved to be most important for objective discrimination between MM and SN: 2c deviation index, 5c exceeding rate, standard deviation of the nuclear DNA content, and both the 85th and the 95th percentiles of DNA distributions. On the basis of these features, we could define a linear discriminant function that allowed a correct diagnosis in 94% of the cases. Our data demonstrate that diagnostically misleading large nuclei in SN are euploid, in contrast to MM. It is thus possible to discriminate SN and MM with high accuracy using DNA cytometry. Because paraffin-embedded tissue can be used, this technique could be a valuable complement to routine histology in equivocal cases.

Bayesian belief networks (BBNs) are a novel tool for representing knowledge about diagnostic decision making and for obtaining a numerical measure of certainty in the final diagnosis. Belief networks have been applied to the pathological assessment of breast, prostate and skin lesions and have been shown to provide consistency in the grading of microscopic features and improve diagnosis. These applications are reviewed in the current paper. The application of BBNs has been further facilitated through the use of standardised imagery which is stored digitally and used to enter evidence into a BBN. It is predicted that the further development of BBNs with improved logical capabilities represent the key to improved decision making in pathology.

Background: Digital image analysis was found to be a useful technique for improved accuracy of preoperative diagnosis of melanocytic lesions. In previous studies digitized color slides were used as input for digital image analysis. New technologies and smaller video cameras made it possible to develop a camera system that allows the digitization of skin lesions directly from the patient., Objective: We investigated whether conventional color slides or directly digitized images should be used for a reliable recognition of malignant melanoma., Methods: Computer features describing characteristics of the lesions were computed for 404 digitized color slides and for 309 directly acquired lesions. Statistical analysis and classifier construction was performed by the commercial statistical classification program CART., Results: With the data set derived either from the color slides or from the directly digitized lesions a sensitivity of about 90% for the recognition of malignant melanoma could be obtained., Conclusion: Both image acquisition techniques allow a reliable detection of malignant melanoma and both are appropriate as input for an image analysis system regarding its efficiency as a diagnostic tool. However, none of the classifiers can be applied with reasonable significance to both techniques.

Recently image analysis (IA) and DNA-cytophotometry (CP) have proved to be useful for the differentiation between benign and malignant melanocytic lesions on paraffin sections. Since, on sections, these procedures are very time-consuming, we tested in the present study whether IA of imprint specimens, which can be evaluated in less than 30 minutes, might also be sufficient. In 39 malignant melanomas (MM), 18 melanocytic nevi (MN), and 6 dysplastic nevi (DN), 12 different morphometric and DNA cytometric features were determined in 100 randomly selected nuclei. In univariate analysis, 5 features were found to be significantly different between the benign and malignant groups (p < 0.0001): mean value (MAREA) and standard deviation (SAREA) of nuclear area and the 80th, 90th, and 95th percentiles of DNA distribution. Using SAREA, the best univariate feature, 82.5% of the cases could be correctly separated. In multivariate analysis with a combination of three features--standard deviation of nuclear area (SAREA), mean DNA value (MDNA), and 95th percentile of DNA distribution (PERC95)--a correct diagnosis was achieved in 89.5% of the cases. Results obtained in the cases of DN indicated an increased proliferation, but did not allow the separation of DN from MM and MN. Since our technique allows a rapid analysis without loss of tissue, which might be important for histological analysis, and the classification rates are equal or still higher than reported in studies on sections, imprints of melanocytic lesions seem to be most appropriate for the calculation of DNA cytometric features as helpful diagnostic criteria in equivocal melanocytic lesions.

The incidence of malignant melanoma, the most dangerous skin cancer, has increased rapidly during the last decade, and the figures are still rising. Since well-trained and experienced dermatologists are able to reach only a diagnostic accuracy of about 75% in visual preoperative classification, the discriminating ability of digital image analysis was evaluated in more than 350 malignant melanoma and benign melanocytic lesions that had all been confirmed histologically. Color slides of melanocytic lesions were scanned and digitized. Computer algorithms were programmed in FORTRAN on a DECstation 5000/200. A feature set was calculated describing the texture, color and their distributions as well as asymmetry, size and border of each lesion. These features, together with the histologic diagnosis, were the input in a commercial statistical classification program. In contrast to the accuracy of 75% achievable by the human eye, a correct classification rate of about 92% was reached with the mathematical classifier as compared with the histologic diagnosis.

Prompted by the well-known difficulties of reliable and objective histologic differentiation between initial malignant melanoma (MM) and benign nevocytic nevi (NN), ultrastructural high-resolution image and multivariate analyses were evaluated for their diagnostic efficiency. Thirty-seven different features describing morphometry (area, circumference, and shape factor), amount of heterochromatin and euchromatin, chromatin homogeneity, and presence of smaller dark chromatin aggregations were determined by a MICROVAX 3500 computer in each of 1840 intraepidermal melanocytic nuclei of 17 MM and 20 NN. A strategy for the classification of cases based on the identification of markedly atypical melanocytic cells (MACS) was developed. MACS, selected in multivariate analysis with a linear combination of the eight most important features for cell classification, were found in 39.4% of the melanoma cells, but only in 0.3% of nevocytic nevus cells. The presence of MACS allowed a clear differentiation between MM and NN. All cases of MM had more than four MACS, whereas 17 cases of nevocytic nevi were MACS negative, and in each of the remaining three cases only one MAC was present. The percentage of MACS detected within intraepidermal parts of MM by using computerized high-resolution image analysis was found to be a highly efficient diagnostic marker. The new classification strategy has the potential of saving considerable time in subsequent studies, because preselected sampling and the calculation of only a few criteria have proven sufficient for correct classification of malignant melanomas.

The histologic and immunohistologic differential diagnosis between pseudolymphomas (PL) and malignant lymphomas (ML) of the skin can be difficult. Since DNA cytometry has been found to be of both diagnostic and prognostic value in various neoplasms, its ability to discriminate between ML and PL in Feulgen-stained imprints of 17 PL and 49 ML skin biopsies was examined by high-resolution image analysis. The reliability of the following algorithms of DNA distribution was evaluated: 1) 2cDI (2c-deviation index), which reflects the variation of the nuclear DNA values around the diploid DNA peak; 2) percentage of cells having a DNA value greater than or equal to 5c (5cER; 5c-exceeding rate); 3) percentage of cells presenting with a DNA value greater than or equal to 4c (4cER). A 2cDI of 0.1 was found to be the most reliable marker for the differentiation between PL and ML. On the basis of this feature, 16 of 17 cases of PL and 46 of 49 cases of ML were correctly classified. The sensitivity, specificity, and efficiency of this feature were 94%. A 5cER greater than or equal to 1% had a specificity of 100%, but the sensitivity was only 43%. For the 4cER, a sensitivity of 61% and a specificity of 94% were found. In conclusion, the calculation of the 2cDI and the 5cER based on high-resolution image analysis provided additional helpful diagnostic features, and therefore should be included as a diagnostic tool. If the 5cER is at least 1%, the diagnosis of a ML can be confirmed with a specificity of 100%.

The first and most important task of automatic high-resolution cell image analysis is the segmentation of the scanned cells. Different methods of cell segmentation have been developed, but a comparison of the capabilities of such algorithms has not been done. This study evaluated two different segmentation methods for cell images, namely, a nonlinear gradient algorithm with a subsequent tracing method and a thresholding algorithm based on the information from three histograms with a subsequent nonlinear cleaning of the binary thresholded images. The same Papanicolaou-stained cell data base was used in both methods. Automatic segmentation of nucleus and cytoplasm was performed, and a comparison with visually segmented areas of the nucleus and cytoplasm was carried out. The difference between the visual method and the automatic segmentation method by thresholding is discussed in terms of classification results.

Human erythropoiesis is an appropriate model for analyzing correlations between functional activity of cells during maturation and nuclear morphology as determined by high-resolution image analysis. In a previous study a relationship had been found between nuclear condensation and the proliferative activity of the cells. This present study analyzed the correlation between nuclear morphology and the protein synthesis rate. The latter was assayed by measuring the 14C-1-leucine incorporation rate utilizing the technique of quantitative 14C-autoradiography. The labeled cells were first classified according to conventional cytological criteria into four groups of increasing maturity. Following grain counting the nuclei were Feulgen-stained, and after removal of the silver grains, the local nuclear optical densities were evaluated by scanning. There was a nonlinear relationship between the protein synthesis rate and features representing the degree of chromatin condensation. This nonlinearity was explained by the mediator function of RNA, predominantly mRNA. The amount of protein produced at a given time depends on the transcriptional activity of the chromatin, the frequency of mitotic divisions partitioning the mRNA, and the half-life of the mRNA. It was concluded that the chromatin texture of erythroblasts reflects three different metabolic activities: the rate of DNA strand duplication, the transcriptional activity for structural proteins enabling a cell to grow and cycle, and for functional proteins, particularly hemoglobin. Since these three activities appear to be synchronized, it is understandable that functional as well as textural features can be used to perform a supervised classification by means of multivariate analysis. Application of the five most significant features allowed a consistent classification of the erythroblasts into the various cytological compartments in 76.7% of all cases, while the classification error of the observer amounted to 13.7%.

The rapid progress in computer technology makes possible the automatic analysis of thousands of cells on a slide in the field of automatic uterine cancer cytology. Our approach starts with the high-resolution scanning of visually selected and classified single cells determining the training set for discriminant analysis. On the basis of more than 15 morphologic and textural features measured, correct classification results of 95% are reached. Our aim is to speed up our image processing system by means of an Array-Processor, the model AP 120 B from Floating Point Systems. In this study a realistic time estimation of the scanning, segmentation, feature extraction and classification of about 40,000 cells on a slide is performed.

Images of exponentially growing mouse L fibroblasts were used to test whether G1, S, and G2 cells could be discriminated by means of densitometric, geometric, textural and chromatin features. The cells were preincubated with FUdR and labeled with 3H-TdR. The definition of G1, S, and G2 was based on autoradiography used to define labeled S cells, and DNA content used to define unlabeled 2c (G1) and 4c (G2) cells. The methanol-fixed Feulgen-stained nuclei were scanned with a TV Plumbicon camera equipped with an array-processor system. Thirty-one nuclear features were calculated and stored together with the precise coordinates of each cell for later relocation and correlation with the autoradiographic data. The features were geometric, densitometric, textural and chromatin parameters of the Feulgen stained images. Feature evaluation and supervised learning were performed. When DNA content was excluded 14 textural and chromatin features were selected. By using these a correct classification of about 80% was achieved. Inclusion of DNA content and its correlates caused degeneration of performance, yielding 75% correct classification. DNA content alone gave 69% correct classification.

Human lymphocytes, transferred in culture into medium conditioned with phytohemagglutinin, were studied with the combined use of quantitative 14C-uridine autoradiography and high-resolution image analysis. During the first 24 h before the onset of DNA synthesis it was possible to study pure G0 and G1 populations. The return of the proliferating cells into the G0 state was clearly monitored in an RNA synthesis rate/DNA content plot during 1-week growth. On days 3-6, the proliferation decreased progressively, and more cells accumulated in the quiescent G0 state. Geometric, densitometric, and textural parameters evaluated from the Feulgen-stained nuclei of the same cells indicated that the decondensation/condensation processes of the nuclear chromatin during the G0 in equilibrium G1 transitions paralleled the RNA synthesis activity. The results help to associate descriptors of nuclear morphology, as derived from image analysis, with functional parameters during activation and return to quiescence of human lymphocytes.

The nuclear morphology of preneoplastic and unaltered hepatocytes in diethylnitrosamine-treated rats was investigated. Two-micrometer-thick sections of methacrylate-embedded liver were scanned with a TV camera and evaluated in a computer using multivariate analysis methods. The preneoplastic cell populations (islands) were distinguished from unaltered hepatocytes by histochemical demonstration of glycogen storage in specimens from starved animals. After the hemalaun-stained liver sections were scanned randomly, the sections were stained for glycogen, and the previously registered cells were identified visually using a scanning stage for relocation. This objective identification of unaltered and preneoplastic hepatocytes formed the basis for the selection of a training set for feature evaluation and supervised classification. Image analysis for quantitative nuclear morphology was applied to the hemalaun-stained cells. The results showed that condensed chromatin was reduced and nuclear area was increased in the nuclei of glycogen-storing cells. Differences in the nuclear structure were also found. A multivariate analysis including seven features gave a correct classification result of about 82%.

Image analysis was performed using Feulgen-stained smears of L cells that were either confluent or had received a nutritional stimulus to proliferate. The cells were grown directly on slides in order to minimize damage to them during manipulation and to avoid trypsin treatment. A series of experiments was performed by varying the dose and the time of the stimulus. The nuclei were interactively sampled and automatically segmented. Most of the information about the fine chromatin structure was gained after subtracting the original image from a median-filtered image. The results indicate that there were changes in the Feulgen-stained nuclear morphology and densitometry shortly after transition of quiescent G0 cells to the proliferating G1 compartment. Multidimensional cluster analysis provided model-independent discrimination between the resting and the stimulated cells; the latter presented increased nuclear dispersion, indicating increased template activity of the cells. The results obtained are applicable both to obtaining a better understanding of the mechanisms of regulation during the G0-G1 transition and to deriving morphologic criteria applicable in diagnostic routines.

Machine recognition of ectocervical cells has been achieved with good classification success by a number of research groups. Even though a substantial number of cell image features have been introduced, only a moderate number are required for classification. For them, definitions and measures of their discriminatory potential are presented.

Cells of a benign pigmented mole and a malignant melanoma were used to compare electron microscopy (EM) and light microscopy (LM) with high-resolution TV-scanning and multivariate analysis methods. Special emphasis was placed on different kinds of chromatin texture features and their discriminating power. Comparative studies on EM and LM are useful to look for the reliability of light microscopic features for the recognition of malignant melanoma cells. Identical cells of benign pigmented moles and malignant melanomas could be scanned by both EM and LM techniques. As a result, there is a significant correlation between EM and LM texture features, derived from the co-occurrence matrix and the difference statistics histogram. This helps to demonstrate that classification in LM is sufficiently good using texture features only.

The segmentation of scenes of fixed tissue sections for quantitative histopathology is the crucial step for further image processing. Different segmentation methods for the separation of nuclei, nucleoli and whole cells in methacrylate-embedded sections of rat liver were investigated. Reasonable segmentation results were obtained using a contrast-enhanced polar-coordinate transformation to distinguish nuclei, a compactness algorithm to distinguish nucleoli and a skeletonization algorithm to delineate cells when a priori information on the morphometric and photometric properties of the liver tissue was included.