Image cytometry is a technique for quantitatively analyzing slide-based, cellular specimens. Images of specimens are acquired using a microscope coupled to an electronic camera and computer, and stored digitally in computer memory, where they are accessible for subsequent analysis by software algorithms.

By first staining the specimen, it is possible to visualize specific molecular species in each cell. In general, the first step in image analysis is the definition of each cell, which is followed by measuring the amount and distribution of stains within individual cells, measuring cell morphology, and measuring the cellular organization of the specimen1. Fluorescent stains have several advantages over using conventional (colorimetric, light-absorbing) stains: (1) non- interfering images of multiple (typically 3-5), spectrally distinct fluorescent labels can be acquired 2; (2) high sensitivity for detection of low concentrations of specific molecules 3, and (3) high accuracy when quantifying the amount of labeled molecules 4. When many (100s) of cells require analysis, which is frequently the case for clinical specimens that contain heterogeneous cell populations, automation of the analysis is necessary 5.

We have developed automated fluorescence image cytometry for analyzing cells and nuclei in cell cultures and clinical specimens 5,6. First, microscopic images of specimens labeled with a fluorescent DNA stain (e.g. DAPI, propidium iodide) were acquired. Since each nucleus contained abundant DNA, then the resulting images had high contrast with bright and dark regions corresponding to nuclei and background respectively. Next algorithms which calculated threshold intensities and employed morphological operations were used for automatic segmentation of each nucleus. These algorithms were tested on images of standard fluorescent micro spheres, cultured cells, cervical PAP smears and tissue sections from prostate and breast tumors. 100% of isolated objects (beads and nuclei) were correctly detected, based on visual comparison of the algorithms' results with the acquired images. However, objects that were clustered together were not always correctly detected, because a dark region was not present between them.

Consequently, we developed analysis algorithms for recognizing clusters and dividing theminto individual objects7. Clusters were recognized by their size and shape and then dividing pathswere calculated to cut them cut into single objects. The dividing paths possessed the highestaverage gradient per pixel out of all reasonable paths across the cluster. The algorithms correctlydetected 94% of the clustered nuclei in 2 mm sections of breast and prostate tumors. The reasons for this high success rate were that the algorithms simultaneously utilized both intensity and shape information available in the images, had low sensitivity to noise and could detect objects with varying sizes and shapes.

After object detection, the amount of each fluorescence label within individual objects may be quantified by integrating pixel intensities within the regions representing each object in images of each label. We measured the quantitative precision of the image acquisition system and analysis software for such measurements by analyzing specimens of standard fluorescent micro spheres and obtained a precision between 2% and 3%4. Similar results have been obtained by other investigators using cellular specimens labeled with fluorescent DNA dyes8. Although fluorescent labels in conjunction with a low light level camera provided high sensitivity, autofluorescence often limited sensitivity in clinical specimens.

We therefore developed a method to correct for autofluorescence. Specimens were imaged using two excitation wavelengths, including one which did not stimulate the fluorescent labels being analyzed. Autofluorescence in the image of the fluorescent labels was calculated as a fraction of the autofluorescence in the image acquired at the non-stimulating wavelength. This calculated autofluorescence was then subtracted from the images containing the fluorescent label of interest. Using this approach, it was possible to detect fluorescence in situ hybridization (FISH) signals of the erbB2 gene in bladder tumor specimens that were invisible in the acquired images 9.

We have applied image cytometry to the quantification of fluorescent labels in cultured cells, cervical PAP smears and breast sections. In the smears, the HPV 16/18 copy number of each cell nucleus (labeled using FISH) was measured and it was found that the HPV status of the specimen positively correlated with the morphological diagnosis10. Other data indicated that E6 (an antigen expressed in HPV 16/18 infected cells and labeled with a fluorescent antibody) was expressed in cervical cells with normal DNA ploidy, suggesting that E6 expression may be an early marker of cervical dysplasia 6. Studies presently focus on the development of analysis software for analyzing 3D images of fluorescence-labeled specimens imaged by confocal microscopy. This is because 3D images are much more precise representations of inherently 3D specimens than 2D images. (The latter only represent an arbitrary slice through the specimen.) For example, FISH signal enumeration in solid tumors is currently performed either on disaggregated cells from the specimen, on touch preparations or on thin (4-6 mm) sections.

Analysis of disaggregated cells suffers from the fact that the cellular organization of the tissue is lost. Touch preparations are biased in favor of those cells which most easily detach from the specimen and are thus not representative of the specimen. Thin sections do not contain intact nuclei thus the number of FISH signals per nucleus may be underestimated. We therefore developed interactive algorithms for enumerating FISH signals in individual, intact nuclei inside thick (20 mm) tissue sections11. Using these algorithms, over 80% of diploid nuclei contained 2 signals from centromeric probes (the expected number) compared to less that 40% when a 4 mm section was used. Automation of these algorithms is underway and preliminary results suggest that 3D analysis will yield improved results over existing 2D algorithms.

In conclusion, automated fluorescence image cytometry is a technology that provides quantitative molecular and morphological analyses at the individual cell level in slide-based specimens. Such information is important for deducing underlying biochemical processes, enables accurate investigation of heterogeneous clinical samples and enables direct correlation of molecular information with cytological and histological diagnoses.

_________

1. Ploem et al (1986). Histochem 84:549-55.

This work was supported by a Biomedical Engineering research grant from the Whitaker Foundation, and by the Director, Office of Energy Research, Office of Basic Energy Sciences, of the U.S. Department of Energy under Contract No. DE-AC03-76SF00098.