Operator input. In the presented method, the input of the operator is reduced to one unavoidable step: specification of the boundary of the studied AOI. The use of an identical AOI′ in the same section for control purposes ensures that changes relative to an indifferent area are calculated in each case even if, by chance, the specification of the AOI is not absolutely precise. If the AOI is to some extent smaller than the area delineated by the boundary of the anatomical region to be analyzed, the area will be under-represented, but the calculated relative difference between the AOI and AOI′ will not be considerably affected. If the AOI is somewhat larger than the area to be analyzed, the measured difference from the control region will be only slightly ‘diluted’ with data from an unaffected area, that is, the method will merely lose some sensitivity. The major landmarks in the sections, however, allow easy and reliable delineation of the area AOI (see Fig. 2). Accordingly, all such inconsistencies would cause relatively minor changes in the results.
Determination of the background staining. In the present experiments, no counterstaining was performed, and there was therefore no need for spectral (colour) aided segmentation of the IHC-stained profiles from the counterstained background, which would have required complex operations (van der Laak et al., 2000; Brey et al., 2003; Pham et al., 2007). Thus, instead of the selection of a specific wavelength (or colour) for analysis, or decoupling of the intensity data from the red–green–blue colour model, all images were converted to greyscale images represented only by 8-bit data at each image point, and the image analysis procedures were performed in a single channel with simplified calculations.
In each section, the intensity of the background staining (determined in an indifferent reference area) is used to identify image points with a level of staining significantly above that of the surroundings both in the AOI′ and within the AOI. Thus, the procedure ensures that the variability of staining in repeated experiments or the heterogeneity of the staining from section to section in a given series will be compensated. The procedure, which automatically calculates the background intensity on the basis of the distribution of the staining intensity without the aid of the operator, consists of two steps. In the first step, an average background image is calculated by applying a large-format low-pass filter to the digital image of the section, which yields a greyscale value at each image point with the average for its 512 × 512 pixel neighbourhood (Figs 3(E) & (F)). Since the great majority of the image points show no evidence of staining (Figs 3(C) & (D)), the result of this filter approximates closely, though not perfectly to the true background distribution (Figs 3(E) & (F)). After low-pass filtering, an average grey value (approximate background level) can be calculated within the AOI′ (233; Fig. 3(F)). The deviation of this number from the true (unknown) background is obviously greater in those regions where stained profiles is notably accumulate (cf. Figs 3(C), (E) and (D), (F)). To correct for this ‘contamination’ of the background distribution from the stained cells, as a second approximation, all image points with grey values lower than the calculated first background value (233) are excluded from the calculation, and a new average background value (BCKGND = 238) is determined for the remainder of the pixels within the AOI′. This procedure of successive approximation of the true average background within the AOI′ should in principle be stopped when the difference between the results of successive iterations is smaller than a specified limit. In the presented experimental example, even in the third step, the change in the calculated average background greyscale value was less than one digit on the 0–255 scale, and accordingly no further iterations were performed.
Segmentation of stained profiles and calculation of parameters characterizing the stained structures. After determination of the BCKGND and the SD values of the greyscale distribution in the AOI′, a cut-off level is determined to segment those image points that are considered significantly stained, that is, pixels with grey values significantly different from that of the background. This seemingly arbitrary decision is supported by statistical principles if the deviation of the cut-off level from the BCKGND value is set in units of SD, which determines the confidence of the results. In biological applications, this offset is normally set to a value in the range 2 × SD to 3 × SD implicating a probability between 67 and 99% for the segmented image to reflect the distribution of the authentic stained profiles. In our application example, the level BCKGND – 2 × SD was consistently applied.
After segmentation, different parameters were derived to characterize the extent and strength of the IHC staining in the AOI relative to the AOI′. The difference between the numbers of pixels segmented in the AOI and the AOI′ (since these areas are identical) reflects the net increase in the stained cellular profiles due to the treatment. For easier comparison of the results obtained from different sections and different animals, these raw data were normalized to the total number of pixels within the AOI or the AOI′, and expressed as the percentile area difference (Δ(area)). The difference in staining intensity between the AOI and the AOI′ was determined directly by using the attenuation (Bouguer–Lambert–Beer) law, applied to the whole area of the AOI and the AOI′. Through measurement of the average (optical) density in these areas, a number corresponding to the difference in the percentile stained area was obtained (Δ(concentration)). The method is analogous to the classical procedures applied successfully in conventional (electron) microscopy to determine the local mass of a specimen by measuring the contrast of the micrographs (Halliday & Quinn, 1960; Zeitler & Bahr, 1965; Edie & Karlsson, 1977).