Differential diagnosis of malignant and reactive cells from serous effusions: image and texture analysis study.

The interrelationship of textural primitives which define morphological texture can be estimated by quite different descriptors; the discriminant value of which varies considerably. In the present study three different approaches to the texture analysis of nuclear chromatin were assayed to correctly allocate 332 cells from a pool obtained from serous effusions (six malignant mesotheliomas, six reactive mesothelial proliferations and five pleural metastases of lung adenocarcinoma). In all cases, initial cytological diagnosis posed considerable problems and final diagnosis was established by histologic examination of surgical specimens. The three approaches were based on binarization of the image obtained by edge detection, gradient analysis and pattern spectrum by morphological opening-closing, respectively. Characteristics afforded by each method were: (a) spatial distribution of heterochromatin, besides geometric features, (b) features related to transitions and contrast between dark and light chromatin primitives, and (c) granulometric characteristics of the theoretically biphasic heterochromatin-euchromatin image defined by mathematical morphology. The three methods were applied to the raw grey-tone image and did not require interactive handling. Although each of the three approaches yielded relatively satisfactory results, with percentages of well-classified cells in the test set ranging from 61.45 to 67.47, the best results (78.31% of well-classified cells) were obtained taking into consideration the three types of variables (area, 2nd opening, 5th closing, and S.D. of the amplitude image). A point to be stressed is the considerably high proportion of correctly-allocated reactive mesothelial cells (82.0%) in a field where subjective assessment commonly yields rather poor results. Nevertheless, classification yielded 14.8% and 3.3% false positives as adenocarcinomatous and malignant mesothelioma cells, respectively. In the theorectical situation devised in the study, results on a cell-by-cell basis encouraging and suggest that a textural approach might be useful in a dedicated expert system or on a more real case-by-case basis.