An Original Approach for Quantification of Blood Vessels on the Whole Tumour Section

Relative abundance of tumour angiogenesis has been shown to be of clinical relevance in cancers of various locations such as the ovary. Nevertheless, several problems are encountered when quantifying tumour microvessels: (i) as many other tumour markers, vascularity pattern is often heterogeneous within the tumour mass and even within the same histological section. As a consequence, an adequate acquisition method must be developed for accurate field sampling. (ii) Manual microvessel counting is long, tedious and subject to poor reproducibility. Introduction in routine practice requires a fast, reproducible and reliable automatic image processing. In this study we present an original procedure combining a slide scanner image acquisition and a fully automatic image analysis sequence. The slide scanner offers the advantage of recording an image of the whole histological section for subsequent automatic blood vessel detection and hot spot area location. Microvessel density and surface fraction were measured for the whole section as well as within hot spots. Different immunostaining methods were tested in order to optimise the procedure. Moreover, the method proposed was submitted to a quality control procedure, with reference to interactive identification of microvessels at scanner level. This experiment showed that 93 to 97% of blood vessels were detected, according to the staining protocol used. Colour figures can be viewed on http://www.esacp.org/acp/2003/25‐2/kim.htm.


Introduction
It is now well established that tumour growth and spread are highly dependent on neo-vascularization [14,32]. A more disputed statement is the prognostic value of tumour angiogenesis. While several studies have demonstrated its predictive significance for clinical outcome in cancers of many locations [1,2,10,12,[17][18][19]30,39], others failed to find any correlation between tumour vascularity and behaviour [3,4,16,25,27,28,36]. These discrepancies could be partly explained by differences in microvessel counting methodology, applied to heterogeneous tumours [5,16]. measured. Usually, and according to Weidner's procedure, the quantification is performed in one or several highly vascularized fields (called "hot spots"), previously selected at low magnification by a trained pathologist [43,46]. Other authors have chosen to systematically sample fields over the whole histological section [22] and to average blood vessel density within the most vascularized fields. However, as these authors noted, the reproducibility in systematic tumour sampling is determined by both tumour heterogeneity degree and sampled tumour area, which is defined by the number of sampled fields. The choice of the microscopical magnification also influences the final tumour area analysed, as well as the reliability of the measured parameters [22,34].
Whether the measures obtained within hot spots or all over the section carry the same biological information remains questionable.
In this context, we propose in this paper a fully automatic method for tumour angiogenesis quantification over the whole histological section and inside automatically detected hot spots. The originality of the study is to assess the degree of neovascularization by processing a single numerical image of the whole tumour section, using a photographic slide scanner, provided with a glass slide holder. A quality control was performed in order to assess the robustness of the quantification method and to find the most suitable immunostaining procedure.

Specimens
Formalin-fixed paraffin-embedded samples of 19 human ovarian tumours were cut into 5 µm thick sections and mounted on polylysine coated slides for immunohistochemical staining.

Immunohistochemistry
Sections were immunolabelled with a polyclonal anti-Von Willebrand Factor antibody (Dako, France), using the Optimax automatic machine (Biogenex) to guarantee staining reproducibility. After deparaffinization and rehydration, slides were heated 30 minutes at 95 • C in a citrate buffer (ChemMate TM , Dako) for antigen retrieval. Following incubations were performed at room temperature. Endogenous peroxidase activity was inhibited by incubating sections for 20 minutes in 0.3% H 2 O 2 . A buffered casein solution (Power Block TM , Biogenex Menarrini, France) was then used to block unspecific binding sites (for 10 minutes). After incubation with the rabbit anti-Von Willebrand Factor antibody (primary antibody, Ab1), 13 and 12 slides were respectively treated by a classical biotin-streptavidin peroxidase technique (series 1) or submitted to signal amplification (series 2) (En Vision TM kit, Dako). Briefly, for classical procedure, after Ab1 incubation (dilution 1 : 50, 1 hour), a biotinylated anti-rabbit antibody (Dako) and then a streptavidinperoxidase complex (Dako) were applied (dilution 1 : 40, 30 minutes for each reagent). For signal amplification, slides were incubated with peroxidase labelled polymer, conjugated to goat anti-rabbit immunoglobulins for 30 minutes. Staining was completed by incubation, either with diamino-benzidine (DAB) for 10 minutes (series 2a), which stains vessels brown, or with amino-ethylcarbazole (AEC) for 15 minutes (series 2b), which highlights them in red (6 samples for each chromogen, on 2 consecutive sections). Light counterstaining was performed during 10 seconds with Gill's haematoxylin [20] for all specimens.

Image acquisition
Images of the whole immunostained sections were captured using a 2700 dpi (1 pixel = 9.4 × 9.4 µm) slide scanner (SprintScan ss35Plus, Polaroid, France) provided with a special medical slide holder (PathScan Enabler TM , Meyer Instruments, Inc., USA) and connected to a personal computer (Pentium III, 450 MHz, 128 Mo). Large images (up to 30 megabytes, for sections up to 8 cm 2 ) were stored as bitmap files. These large images of the whole tumour tissue section were directly processed for automatic vessel detection.

Image processing
Image processing was performed by chaining operators of an image analysis toolbox Aphelion software, v. 3.0 (Adcis, France) implemented on a personal computer and running under Windows operating system.
The procedure can be divided into 5 steps: background correction, tissue detection and necrosis elimination, immunostained structure detection, hot spot identification and location, and measures (Figs 1, 4, Table 1).
Before any analysis, the user can discard normal tissue surrounding the tumour, by drawing regions of interest.  Table 1 Measures obtained by the automatic procedure (case presented in Fig. 3). All microvessel profile density (MVD) were expressed as number per mm 2 of tumour tissue. All surface proportions were given as a percentage

Background correction
The medical slide holder used for scanner acquisition is provided with a 25% transmission filter, which generates a light blue background over the whole image. In order to restore true colours, a correction was applied by selecting a small area, outside the tissue, and setting its grey values to the maximum (255) in the 3 colour channels (Red, Green, and Blue). The 3 RGB values of each pixel of the image were then computed with reference to these new background values.

Tissue detection and necrosis elimination
A fixed threshold, set on the green component of the original image, allowed the detection of the histological section limits (TSnec + ). Tumour tissue can contain necrotic areas and mucus which might be deeply stained by chromogens and could appear as large red or brown fields. They were easily removed by thresholding an "excess blue image", a grey scale image generated from the linear combination of the 3 components RGB [50] and computed as follows: A mathematical morphology closing by reconstruction was set on this grey scale ExcB image. This operator combines a dilatation (square element of size 10) with a geodesic reconstruction (in a 4-connectivity mode 1 ). It is used in the same situations as the standard morphological closing, but has the advantage of 1 In a 4-connectivity mode, only pixels located in the north, south, east and west of a given pixel are considered as neighbour. In a square grid like the one used in this study, these pixels are the closest to the central one.
preserving shape contours. Thereafter, a threshold was set (minimum grey value; 25) and an opening by reconstruction of size 3 was applied on the resulting binary image, to specifically remove small structures, while preserving these large areas. A logical difference between the last image (Nec) and the image of tumour tissue section allowed necrosis elimination (TSnec − ).

Immunostained structure detection
Detection of highlighted vessels was performed on 2 different images: the first corresponding to the blue channel and the second image (TC) to the following combination: Firstly, a black top hat transform (difference between closed image and initial image) [35] was applied on the blue image and a fixed threshold (8; 25) was set in order to isolate well-contrasted structures. Structures with holes were filled (hole-fill operator, 4-connectivity mode) and elements cut by the image border were removed (border kill operator, 4-connectivity mode) (St).
Secondly, the brown or red staining deposits were detected by automatically thresholding the TC image (entropy threshold according to Kapur et al. [29]) (Mk).
Objects isolated in this last step (Mk) were used as markers for reconstruction of contrasted structures already characterised (St) (4-connectivity mode). Vessels were reconstructed (Vs) and most false-positive structures were removed using this method (background and unspecific stain deposits). Only elements included in the tissue section were kept by a logical operation (logical AND) between TSnec − and Vs (Res).

Hot spot location
Finally, since the whole histological section was analysed, hot spot areas were identified. For this purpose, neighbouring microvessels were aggregated by closing (square element of size 10) which generates regions of various sizes, according to the number of microvessels joined; small areas ( 900 µm 2 ) were removed using a 10-size reconstruction by opening (8-connectivity).

Measures
For each image, mean microvessel profile numerical density (MVD, number per surface unit of tumour tissue) and mean microvessel surface proportions (MSP) were determined. Mean and maximum values of MVD and MSP inside identified hot spots were also computed. Finally, the area fraction of hot spots was estimated.

Quality control procedure
Automatic image processing results in the detection of positive structures (P) as well as false-positive objects (FP). Their respective percentages (%P, %FP) and the percentage of well-detected vessels (%VD) were determined with reference to an interactive identification of vessels.
For this purpose, PaintShop Pro7 (Jasc Software) tools were used for drawing and Aphelion software for computation. Firstly, all vessel profiles observed on the histological section were pointed out, generating a vessel-mask (Vm), caught from the initial image overlay (Fig. 2). The number of dots pointed out on this mask was computed and kept as a reference (nVm).
As mentioned above, the resulting binary image (Res) of the automatic procedure contained welldetected vessels as well as false-positive structures. In order to estimate the proportion of each of them, Res was superimposing on the initial image. By transparency, well-detected vessels could be distinguished from false-positive structures: each structure identified by the observer as a microvessel was interactively pointed out, generating a binary mask called VDm. Structures of Res were then reconstructed from dot markers of VDm. This generated an image of vessels, correctly recognised by the automatic procedures (VD). Image of false-positive structures (FP) was obtained by a logical difference between Res and VD. Interactive pointing was performed in triplicate for each mask (Vm and VDm).
The number of structures present in Res, VD and FP were computed (respectively nRes, nVD and nFP) and the percentage of automatically detected vessel (%VD), positive (%P) and false-positive (%FP) rates were respectively defined as follows:  One can notice that: Surface fraction of both correctly detected vessels (MSP VD ) and false-positive structures (SP FP ) were also determined. The relative surface proportion (rSP) of false-positive vs. correctly detected vessels was defined as follows: Furthermore, a quality factor was computed [9] as follows: the higher the Qt, the better the result (Qt max = 100).
This control procedure was used to evaluate the quality of automatic microvessel detection and to compare results obtained by the various staining procedures tested: (ii) Highlighting blood vessels in brown (DAB as a chromogen) and in red (AEC as a chromogen).

Automatic microvessel detection from scanner images
The filter of the medical slide holder provides a uniform blue background over the whole image. Profiles obtained in the three channels along a diagonal line, drawn across the image show that this background is at the same level all over the image; it can then be easily removed by a fixed threshold (Fig. 3). Figure 4 shows an example of automatic microvessel extraction from a scanner image of a whole histologic section. It takes nearly 15 minutes to process such a 30-Mb image. An interactive delineation of the tumour was first set for removing normal tissue (Fig. 4A). All the following steps were automatically done: tumour tissue contour detection, necrotic areas and mucus removing, vessel detection and hot spot identification.
Microvessel profiles extracted in this example represent 4.5% of the tumour section area; mean microvessel profile density (MVD) inside the entire tumour section was of 40 vessel profiles/mm 2 (Table 1).
Many hot spots were located and visualised on the whole image. These highly vascularized areas represent 17% of the tumour section. The mean and the maximum MVD values inside these hot spots were respectively of 144 and 387 vessel profiles/mm 2 . The mean surface proportion of these microvessel profiles (MSP) inside the hot spots was 19% and its maximum was 30%.

Quality control of the automatic detection
Compared to interactive vessel pointing, most vessel profiles were detected (mean %VD: 97%) after classical immunolabelling and DAB staining, with a mean %FP of 22% (range 14-37%) (Fig. 5). Images with a %FP beyond 30% (T4 and T7) correspond to unsatisfactory stained sections (background and reagent deposits). The mean quality factor Qt obtained reached 75% for the 13 cases studied (range 63-83%).

Method robustness with regard to the staining procedure used
Results of microvessel automatic detection on images of tumour section, immunostained by classical and amplified method (EnVison kit) are given in Fig. 6 (DAB as a chromogen for both staining method). For an equivalent mean %VD (97%), %FP was higher (mean %FP: 30%) and reached 51% in   Comparing results obtained on images of two consecutive sections of the same tumour, immunostained by DAB or AEC, showed (Fig. 7) that mean %VD was lower when using AEC (93%) than DAB (97%) but less false-positives were obtained with AEC, since mean %FP was of 26%, against 51% with DAB. This resulted in a better mean Qt with AEC (69%, range 52-85%) than with DAB (64%, range 45-82%).
Furthermore, the analysis of the relative surface proportion of false-positive structures vs. detected vessels (rSP) showed that automatic image processing isolated false-positives of relatively small size (mean rMSP = 11.85%, range 6.57-23.19%) when vessels of tumour sections were classically stained with DAB (Fig. 8). In contrast, mean rSP was 21.45%  (range 8.95-41.31%) and 19.42% respectively (range 9.91-34.06%) when the amplification protocol was used, either with DAB or AEC, showing that larger false-positives were detected.

Discussion
Colour figures can be viewed on http://www.esacp. org/acp/2003/25-2/kim.htm. Results presented here underline the advantages of an original image acquisition method, coupled to an automatic image analysis procedure, for angiogenesis quantification on immunostained tumour sections. The slide scanner of-fers the opportunity of obtaining an image of the whole histological section, which can be processed in one step. The observer can also discard normal tissue from analysis by selecting a region of interest inside this image. Reproducibility in vessel detection and hot spot identification is possible using this strategy. Furthermore, this approach provides a solution to the major problem of tumour tissue heterogeneity. Indeed, discrepancies found in literature can be partly explained by the bias introduced by human intervention in field selection, hot spot identification or vessel counting [7,46,49], amplified by the well-known problem of biological heterogeneity of tumours (and in particular, intra-tumoural vascular heterogeneity) [5,16].
In such a context, two major points have to be solved: the choice of the parameter to be measured and the measurement method. Concerning the first point, whether the measurement within the most vascularized fields (called "hot spots") is more significant than an average value obtained over the whole section, remains uncertain. Most authors assess tumour angiogenesis by estimating microvessel numerical density (number of profiles per unit area), within hot spots [15,21,33,46]. In these conditions, it is impossible to distinguish uniformly from focally well-vascularized tumours, whose behaviour may be quite different. Only confrontation or combination of data obtained from the whole tissue and from hot spots could give a reliable answer, considering moreover that the presence of such highly vascularized fields is not always obvious. In this manner, Kato et al. showed that the average of microvessel density, but not the highest value, was an independent prognostic indicator in breast cancer [30,31]. The procedure proposed here has the advantage of providing both values, together with the mean value within all hot spots identified on the section. Furthermore, an estimate of the area fraction of hot spots is given.
As vascular density obtained using such an automatic image analysis method may be biased by imperfect segmentation of microvessels (which can aggregate neighbouring profiles or can split a single one into several unconnected fragments), vascular area fraction is given in addition and should be preferred. One must notice that some authors have already suggested that vascular area fraction should also be considered as an important parameter [15,38,47]. Which of these parameters have the strongest clinical information will be determined by studying large series of patients.
The manner in which the fields to be measured are chosen is a very critical step. To give a mean value over the whole section at microscopical level, it is necessary to acquire all fields or to systematically sample the histological section and take into account the mean value. The first solution is time consuming and needs large image storage capacities. The second method is more attractive, but, as previously mentioned, the reproducibility in systematic tumour sampling is determined by both tumour heterogeneity degree and sampled tumour area, which is defined by the number of sampled fields [22]. Acquisition by a slide scanner solved this sampling problem as the histological section is analysed as a whole from a single image (of at most 30 Mb).
If the quantification must be assessed within hot spots, their identification has to be as objective as possible, to ensure reproducibility of data. For this purpose, several authors proposed to sample and analyse the whole section at microscopical level. Vascular parameters are automatically assessed for each field and hot spots are then identified [7,41]. For example, Van der Laak et al. averaged the three highest values obtained in separate fields over the whole section [41]. However, since these fields can be spread out across the histological section, this could not be considered as a true hot spot. To the best of our knowledge, only a recent study proposed an objective identification of hot spots covering more than one microscopical field [7]. The authors presented an automatic blood vessel detection and mapping over the whole section; afterwards, the authors automatically identified hot spot areas from this map. However, as noted by authors, the "automatically" identified hot spots did not always fit with those identified visually, indicating that interactive selection of hot spots, as usually performed, may be biased. Furthermore, the study in question underlined that results obtained were dependent on the shape of hot spots, defined by the observer. In the same way, the hot spot identification procedure presented in this paper is dependent on the size of the structuring element used, which defines the distance between microvessels to be aggregated. In the present study, hot spot areas were determined by joining microvessels distant from at most 200 µm. This distance was determined considering several data available in the literature: (i) the distance between perfused vessel and hypoxic regions in tumours varied from 50 to 150 µm (mean 100 µm) [24,37,42,45,48]. Beyond this distance from the nearest blood vessel, cells die due to the lack of oxygen and nutrient supply, and undergo necrosis [8,37,48]. So, the maximum inter-vessel distance may be nearly 200 µm.
(ii) In invasive breast carcinoma, Belien et al. found a majority of mitotic figures between 0 and 102-268 µm, and a sharp decrease of them beyond a distance of 100 µm from the nearest blood vessel [8]. Then, aggregating vessels separated by a distance lower than or equal to 200 µm is in agreement with previously published data.
Another point to be discussed is the 3D sampling. One must notice that, although removing subjectivity in 2D section sampling, the results obtained here do not take into account 3D-tumour heterogeneity. They were collected from a single slide chosen as "the most representative" of the tumour mass. This "one slide" analysis has some chances to be successfully introduced in routine practice. Nevertheless, the simplicity, low cost and rapidity of the proposed method allow serial section analysis. 3D sampling could be solved if the investigator applies stereological rules; some equipment, such as the tissue slicer, may be helpful for assessing this manipulation [26].
Once the sampling strategy is chosen, a reproducible quantification procedure has to be used. Only a fully automatic system can ensure this reproducibility. Nevertheless, the quality of such a procedure is highly dependent on specimen preparation (especially on tissue staining) [47], and on the resolution of the acquisition material used.
Concerning the first point, recent papers show that signal amplification with biotinylated tyramine improved vessel staining, with poor background noise [7,22]. The quality control of the presented method showed that the signal amplification used generated noise thus giving rise to false-positive events. Automatic quantification procedures found in literature, allow the user to interactively remove false-positive structures; the method proposed in our paper discards any human intervention, but a percentage of microvessels detected as well as an estimate of false-positive objects were evaluated. In addition, it has been reported that the investigator's experience influences the result obtained [6,44]. A fully automatic system does not require extensive training. Furthermore, to be adapted to clinical routine, a method for biological marker quantification ought to be simple and fast. Fully automatic procedures have been already noted to be faster than semi-automatic or manual ones [7,47]. The procedure we propose takes only 15 minutes to analyse a section of approximately 5 cm 2 , as opposed to previously published time consuming methods [7].
Concerning the resolution of the acquisition system, it must be high enough to ensure catching all structures of interest. The quality control performed in the present study showed that most microvessels, interactively identified, were detected (93 to 97%, according to a staining with AEC or DAB). However, it does not guarantee that all microvessels present on the tumour section were detectable by the human eye at scanner level. To answer this question, the procedure developed should be more extensively compared to classical microscopic image processing in order to appraise the possibly lost vessels (work in progress). Moreover, the prognostic value of the vascular parameters provided at scanner level is now being evaluated through a multivariate analysis of clinical and biological markers from a large series of ovarian tumours.