Identification of a characteristic vascular belt zone in human colorectal cancer

Blood vessels in cancer Intra-tumoral blood vessels are of supreme importance for tumor growth, metastasis and therapy. Yet, little is known about spatial distribution patterns of these vessels. Most experimental or theoretical tumor models implicitly assume that blood vessels are equally abundant in different parts of the tumor, which has far-reaching implications for chemotherapy and tumor metabolism. In contrast, based on histological observations, we hypothesized that blood vessels follow specific spatial distribution patterns in colorectal cancer tissue. We developed and applied a novel computational approach to identify spatial patterns of angiogenesis in histological whole-slide images of human colorectal cancer. A characteristic spatial pattern of blood vessels in colorectal cancer In 33 of 34 (97%) colorectal cancer primary tumors blood vessels were significantly aggregated in a sharply limited belt-like zone at the interface of tumor tissue to the intestinal lumen. In contrast, in 11 of 11 (100%) colorectal cancer liver metastases, a similar hypervascularized zone could be found at the boundary to surrounding liver tissue. Also, in an independent validation cohort, we found this vascular belt zone: 22 of 23 (96%) samples of primary tumors and 15 of 16 (94%) samples of liver metastases exhibited the above-mentioned spatial distribution. Summary and implications We report consistent spatial patterns of tumor vascularization that may have far-reaching implications for models of drug distribution, tumor metabolism and tumor growth: luminal hypervascularization in colorectal cancer primary tumors is a previously overlooked feature of cancer tissue. In colorectal cancer liver metastases, we describe a corresponding pattern at the invasive margin. These findings add another puzzle piece to the complex concept of tumor heterogeneity.


A characteristic spatial pattern of blood vessels in colorectal cancer
In 33 of 34 (97%) colorectal cancer primary tumors blood vessels were significantly aggregated in a sharply limited belt-like zone at the interface of tumor tissue to the intestinal lumen. In contrast, in 11 of 11 (100%) colorectal cancer liver metastases, a similar hypervascularized zone could be found at the boundary to surrounding liver tissue. Also, in an independent validation cohort, we found this vascular belt zone: 22 of 23 (96%) samples of primary tumors and 15 of 16 (94%) samples of liver metastases exhibited the above-mentioned spatial distribution.

Summary and implications
We report consistent spatial patterns of tumor vascularization that may have far-reaching implications for models of drug distribution, tumor metabolism and tumor growth: luminal hypervascularization in colorectal cancer primary tumors is a previously overlooked feature of cancer tissue. In colorectal cancer liver metastases, we describe a corresponding pattern PLOS  Introduction

Spatial heterogeneity in solid tumors
Human solid tumors have a high degree of spatial heterogeneity, both on a genetic [1] and on a tissue level [2,3]. This tumor heterogeneity is reflected by non-uniform distribution of microscopic structures of interest within a tumor; for example blood vessels [4], proliferating tumor cells [5] or lymphocytes in primary tumors [6,7] and metastases [8,9]. In general, heterogeneity can be appreciated in many scientific disciplines dealing with spatial data (e.g. earth sciences [10] or radiology [11,12]). However, in cancer research and histological cancer diagnostics, this spatial heterogeneity is often neglected because of methodical limitations. For example, blood vessels and proliferating tumor cells are only counted in selected areas because quantification methods often rely on tedious manual procedures [13]. With the advent of digital pathology, microscopic objects in tumor tissue can be analyzed in a high-throughput manner in the entire spatial domain of a whole slide image (WSI) [2,14]. Consequently, objects (such as blood vessels) and their features (e.g. size or shape) can be measured and subsequently visualized e.g. as a density map (heat map). Within these maps, one can easily appreciate the heterogeneity of the underlying data. Especially, "hotspot" regions of high density can be identified. However, for many applications in histopathology, there is no clear-cut definition of these hotspot areas. In this light, we have recently proposed a new method to analyze blood vessels that is based on spatial statistics and identifies all hotspot areas that are unlikely to occur by chance [4].

Blood vessel heterogeneity in tumors
As tumors grow, they induce the sprouting of blood vessels into the tumor tissue [15][16][17]. Spatial distribution of vessels (blood or lymphatic) has been investigated in some types of cancer [18][19][20]. Also, morphological parameters of intra-tumoral blood vessels have been linked to patient outcome in CRC and other types of cancer [21][22][23][24]. However, to our knowledge, there have been no systematic investigations on the spatial distribution of blood vessels in whole slide images (WSI) of CRC tissue except for our recent work on tumor angiogenic hotspots [4].
In this light, we investigated whether there are other characteristic distribution patterns of blood vessels in CRC tissue besides angiogenic hotspots.

Ethics statement
For this study, we retrieved randomly selected tumor tissue samples from N = 100 patients from our institution's pathology archive (S1 Fig). Prior to statistical analysis, we retrieved the TNM stage from the original pathology report (S1 Table). No other patient data was used within this study. For all analyses, the samples were fully anonymized. All experiments were approved by the institutional ethics board (medical ethics board II, University Medical Center Mannheim, Heidelberg University, Germany; approval 2015-868R-MA). The institutional ethics board waived the need for informed consent for this retrospective analysis of anonymized samples. All experiments were carried out in accordance with the Declaration of Helsinki.

Data availability
We release all raw data (images and measurement values) under a Creative Commons Attribution 4.0 License (http://creativecommons.org/licenses/by/4.0/). The data can be accessed via the following DOI: 10.17605/OSF.IO/D8AR6. This includes raw TIFF images of all N = 100 CRC samples and TIFF images of N = 5 normal colon mucosa. Furthermore, we release N = 100 histological images that were used for validation of the segmentation procedure by human observers: DOI: 10.5281/zenodo.117997. This dataset includes the counts of three blinded observers and can be used by other groups to validate their own blood vessel segmentation algorithms.
Source codes used for this study are available under the MIT license (http://opensource. org/licenses/MIT) and can be accessed via the following DOI: 10.5281/zenodo.260139.

Sample collective
In this study, we analyzed two independent collectives of human tumor samples. We retrieved these tumor samples from our local pathology archive at University Medical Center Mannheim. In our archive search tool, we filtered samples by the keywords "colon carcinoma", "rectum carcinoma", "colectomy" and "hemicolectomy" and reviewed all search results manually. At this stage, we identified pathology reports that described a immunohistologically diagnosed microsatellite instability (MSI), a largely necrotic tumor or a tumor located in the appendix and omitted the corresponding samples. No further selection criteria were applied.
First cohort: We retrieved 65 unselected samples of CRC tissue from the local pathology archive of which four samples were excluded because of insufficient tissue quality, leaving N = 61 samples (see S1 Fig and S1 Table). From each patient, one tumor block was selected based on pre-existing hematoxylin & eosin (H&E) slides and was stained for CD34. One metastatic sample matched a primary tumor, i.e. was derived from the same patient (Smp006 and Smp029).
Validation cohort: The validation cohort was composed of N = 39 randomly selected CRC tumor samples, of which N = 23 were primary tumors and N = 16 were liver metastases. Two metastatic samples matched a primary (C2-Smp025 matched C2-Smp009 and C2-Smp028 matched C2-Smp014). To ensure homogeneity of this cohort, no neoadjuvantly pretreated tumors were included. Representative images from the validation cohort can be seen in S2 Fig. Immunostaining and slide scanning Immunohistochemistry (IHC) was performed using a routine immunoperoxidase technique [25,26] with anti-CD34 antibody (Immunotech PN IM0786, 1:500), and with DAB (Diaminobenzidine) as chromogen. Subsequently, the slides were fully digitalized using an Aperio ScanScope (Aperio / Leica Biosystems, 20x objective magnification, approx. 0.5 μm per pixel) and saved as compressed Aperio SVS files, typically yielding 300 MB per slide. Whole slide images were reviewed and manually cropped, yielding TIFF images of 1-2 GB per sample (uncompressed).

Image analysis and object recognition
Intra-tumoral blood vessels were detected according to our previously published computational method [4], including color deconvolution [27], unsupervised thresholding [28] and morphological post-processing. To optimize computational performance, objects in the binary mask were processed using a dilation-erosion step without further complex morphological operations. On average, blood vessels had a major axis length of 12.7 μm ± 1.7 μm. The blood vessel detection method is explained in our previous publication [4].
The vessel segmentations method returned the centroid coordinates of recognized objects. These coordinates were passed on to our algorithm for automatic detection of angiogenic hotspots [4]. This hotspot detection is based on unsupervised kernel density estimation (KDE) [29]. Particularly, our previous publication includes a description of how the optimal threshold of significance for an angiogenic hotspot is calculated [4].

Analysis of spatial distribution patterns
To test whether objects in a histological whole slide image (WSI) were preferentially located close to the intestinal lumen or the invasion front, we devised and implemented a new spatial statistical approach. After object recognition, we manually delineated up to three ROIs in the WSI: "Tumor", "Adjacent Tissue" and "Intestinal Lumen". For each object in the "Tumor" ROI, we calculated the shortest distance to either of the two neighboring ROIs ("Adjacent Tissue" and "Intestinal Lumen"). For each image, this distribution of distances D obs (x) was compared to 100 control distributions D i ctrl ðxÞ (x = distance in mm, i ¼ fiN j 1 i 100g). These control distributions followed complete spatial randomness (CSR). Blood vessel CSR patterns were created per a 2D Poisson process.
By comparing the observed pattern to the control pattern, we defined a spatial excess E of objects with respect to the distance from a region of interest (ROI) as follows: E x ð Þ ¼ P i D obs ðxÞ À D i ctrl ðxÞ n (n being the number of independent repetitions of the experiment, default 100; x being the distance from "Lumen" or "Adjacent Tissue" ROIs in mm; i being the index variable for iterations from 1 to n). In this study, we use the term "excess" to denote an aggregation of objects of interest (e.g. blood vessels) in a specific region that statistically outnumbers the aggregation of these objects that can be expected under the null hypothesis of CSR.
Since the observed pattern and the control pattern contained an equal number of points, it follows that the distance distribution (visualized as a histogram) contained an equal number of object counts, so that If the objects were aggregated close to the ROI, E(x) showed a positive peak close to x = 0 (seen in Fig 1B). If the objects were sparse close to the ROI, E(x) had negative values close to x = 0. We identified the first x-intersection x 0 of E(x) so that A as defined as EðxÞdx was the absolute object excess close to the ROI (Fig 1B). In other words, E(x) was analyzed for the presence of a positive or negative object excess close to the ROI of interest. An object excess was considered statistically significant if the 95% confidence interval for the true mean did not contain zero. For a more detailed graphical explanation of the method, see S3 Fig. Furthermore, we measured global descriptors of blood vessel distribution patterns as follows: in accordance with Vermeulen et al. [13], we counted blood vessels within angiogenic hotspots ("in-hotspot microvascular density (MVD)"). Additionally, we used the total number of blood vessels within the tumor ROI, subsequently denoted as "overall MVD". Also, the mean number of spatially separated hotspot areas per area was defined as "hotspot density".

Comparison to normal colon mucosa
For qualitative comparison of CRC vascularization patterns to normal colon mucosa, we randomly chose N = 5 anonymized patients from our sample collective and retrieved corresponding non-malignant colon tissue from our pathology archive.

Statistics and computational methods
All values are given as mean ± standard deviation except if otherwise noted. Uncertainty bars in waterfall plots denote the 95% confidence interval of the mean and uncertainty bars in other  (1). The red curve peaks at approx. 0.2 mm. This peak corresponds to an accumulation of blood vessels close to the intestinal lumen. We then assessed whether this peak is due to chance or due to a non-random effect. To this end, a random point pattern was simulated and the experiment was repeated with these random points. This was repeated 100 times. The results of these experiments are plotted in black (3), the mean distance histogram of the random points is shown in blue (2). The peak of the observed curve (1) is far outside the range of random fluctuations. This shows that the spatial accumulation of blood vessels close to the intestinal lumen is very likely not due to a random effect. (B) To quantify the blood vessel excess, the difference of (1) and (2) is plotted as a histogram. The area under the curve until the first x-intersection is regarded as the blood vessel excess close to the intestinal lumen. This Fig shows representative data for one sample (C2-Smp008).
doi:10.1371/journal.pone.0171378.g001 plots denote standard deviation (see also the respective Fig legends). To test for significance, we used a two-tailed student's t-test with a level of significance of 0.05, except if otherwise noted. All computational methods in this study, e.g. image processing methods, spatial statistical calculations and data analysis were performed by self-developed MATLAB programs (MATLAB, Mathworks, Natick, MA, USA). For our experiments, we used a standard computer workstation (2.2 GHz Intel Core i7, 16 GB 495 DDR3 RAM). To analyze a whole slide image typically took between 20 and 40 minutes. No systematic effort was made to further optimize for computational performance. Further computational parameters are available in S1 List.

Automatic blood vessel detection is equivalent to manual detection
In the present study, blood vessels were automatically detected in histopathological images. We validated our segmentation method by comparing automatic blood vessel counts in N = 100 histological image patches against three blinded human observers. The results of this experiment can be found in Fig 2. Here, it is evident that blood vessel counts by human observers show a certain inter-observer-variability (Pearson's correlation coefficient of observer (Obs.) 1 vs. Obs. 2: r = 0.98; Obs. 1 vs. Obs. 3: r = 0.93; Obs. 2 vs. Obs. 3: r = 0.95). Computer-based object recognition correlates well to the mean of human observers (r = 0.81). Although this correlation value is below the correlation between different human observers, a correlation of r = 0.81 generally indicates a strong positive relationship between two variables. We conclude that for the purposes of our study, the blood vessel segmentation method is sufficient.

Blood vessels in CRC primary tumors form an angiogenic zone close to the intestinal lumen
By visually inspecting continuous maps of intra-tumoral blood vessel density in CRC samples, we saw that angiogenic hotspots were accumulated in an area of the tumor facing the intestinal lumen in CRC primary tumors (Fig 3A and 3C). Conversely, in CRC metastases, we saw that angiogenic hotspots were predominantly located at the invasion front of the tumor (Fig 3B  and 3D). Histologically, it was evident that these automatically detected hotspots align with tightly packed aggregates of dilated blood vessels (Fig 4C, S4 Fig) in primary tumors and aggregates of small vessels in metastases (Fig 4B). These aggregates were not present at the invasion front of primary tumors (Fig 4A). This observation led to our main hypothesis: "Blood vessels abundantly occur in a spatially distinct, non-random, vascular belt zone that is located close to the intestinal lumen in primary tumors and close to the invasion front in liver metastases". The next step was to test this hypothesis, objectively quantify these subjective observations and to assess the statistical significance of the observed pattern.
We developed and employed a statistical approach to find non-random spatial distribution patterns of intra-tumoral blood vessels in colorectal cancer samples. Initially, we automatically analyzed more than 300,000 vessels in N = 61 tumor tissue samples (50 primary tumors, 11 metastases; this is referred to as the first cohort). The intestinal lumen could be identified in N = 38 WSI of CRC primary tumor samples (of which N = 4 had undergone neoadjuvant treatment and N = 34 had not). In these N = 38 samples, we quantified the distribution of distances of blood vessels to the intestinal lumen. We found that in 33 out of 34 untreated CRC tumor samples (97%), blood vessels were significantly more numerous at the luminal side than would have been expected under a random spatial distribution (mean excess 457 ± 355 blood vessels, Fig 5A). The positive blood vessel excess was indeed restricted to a clearly defined zone of 1.27 ± 0.40 mm adjacent to the intestinal lumen. In pretreated CRC tumors, the pattern was different and a positive blood vessel excess at the luminal side was only observed in 2 of 4 samples and to a much lesser degree (mean excess 147 ± 265 blood vessels, width of the positive excess zone 1.02 ± 0.26 mm, Fig 5A)-although, because of the limited sample size of pretreated (neoadjuvant) CRC tumors, these numbers have to be viewed with caution. To exclude the possibility that the observed pattern is merely an artifact present in our sample collective, we assembled, stained and analyzed a second sample collective (referred to as validation cohort) of N = 39 human tumor samples (N = 16 metastases and N = 23 primary tumors). An overview of all samples is given in S1 Fig. A quantitative analysis of this cohort (Fig  6) completely reproduced the findings from the first cohort (Fig 5): 22 of 23 CRC primary tumors (96%) showed a significant positive blood vessel aggregation close to the intestinal lumen.
Comparably to the samples in the first cohort, primary CRC tumor samples in the validation cohort exhibited a highly vascularized zone of mean width of 1.11 ± 0.37 mm. Mean blood vessel excess (as compared to a random spatial distribution) was 579 ± 450 blood vessels. These findings from the validation cohort are in qualitative and quantitative agreement with the first cohort.
Having analyzed the distribution of blood vessels relative to the intestinal lumen, we analyzed the distribution of blood vessels relative to the adjacent tissue (invasive margin). This is described in the following paragraph.

Blood vessels in CRC liver metastases cluster in an angiogenic zone at the tumor invasion front
In the first cohort, the tumor invasion front was present in N = 51 of 61 histological whole slide images (among these: N = 36 untreated primary CRCs, N = 4 neoadjuvant CRCs, N = 11 CRC liver metastases, Fig 5B). We analyzed the distance of blood vessels to the adjacent tissue and found a marked positive blood vessel excess at the invasion front in all 11 of 11 metastases (mean excess 440 ± 198 blood vessels, Fig 5B). The mean thickness of this blood vessel excess zone was 1.36 ± 0.37 mm. In untreated primary CRCs, we observed a significant positive blood vessel excess at the invasion front in only 12 of 36 samples (33%). In these 12 samples, the mean excess in the positive excess zone was 156 ± 131 vessels and the mean thickness of the positive excess zone was 0.78 ± 0.37 mm. In short, while a vascular belt zone close to the adjacent tissue was observed in all metastases, only 33% of untreated primary CRCs exhibited a positive blood vessel excess zone and this zone was much thinner and less pronounced than in CRC metastases (Fig 5B and Fig 6B).
Like before, we validated these findings in our validation cohort. The tumor invasion front was present in all but one (38 of 39) sample in this cohort (N = 23 primary tumors and N = 16 metastases). As reflected in the primary cohort only a few (3 of 22, 14%) primary tumors exhibited a blood vessel excess close to the invasion front. Conversely, 15 of 16 (94%) of all liver metastases in the validation cohort showed this specific pattern. Thus, the findings in the validation cohort completely reproduced the findings from the first cohort. In summary (as sketched in Fig 7), liver metastases and primary tumors exhibited quite different spatial distribution patterns of blood vessels close to the invasive margin. This difference in "adjacent tissue blood vessel excess" was highly unlikely a product of chance (p < 0.001).

Different angiogenic activity in metastases than in primary tumors
In the present study, we could replicate our previous finding that angiogenic hotspots are spatially clearly defined, contiguous areas in a tumor image [4]. We quantified the size and number of these hotspot areas in all samples and found pronounced differences between primary tumors and metastases (given for the combined data set consisting of both cohorts, excluding neoadjuvantly pretreated tumors): In liver metastases, hotspot density was much higher while overall microvascular density (MVD) was markedly lower than in primary tumors: Hotspot density was 0.37 ± 0.13 hotspots/mm 2 in untreated primary tumors (N = 68) and 0.44 ± 0.17 in metastases (N = 27), with p = 0.026. MVD was 84 ± 22 vessels/mm 2 in primaries (N = 68) and 74 ± 19 vessels/mm 2 in metastases (N = 27), with p = 0.046.

Comparison of total microvascular density, in-hotspot microvascular density and number of hotspots
In-hotspot MVD is a widely-used parameter to assess tumor vascularization [13], but it is unclear how it correlates to overall MVD. We analyzed both parameters in N = 100 samples (first cohort and validation cohort) and found that in metastases and primary tumors, overall MVD and inhotspot MVD are well correlated (Pearson's correlation coefficient r = 0.67, S5 Fig).

Characteristic angiogenic patterns in CRC tissue
In the present study, we investigated spatial patterns of blood vessel distribution in human CRC samples for the first time and found characteristic patterns: (1) intra-tumoral blood vessels were aggregated in a clearly defined zone close to the intestinal lumen in primary tumors.
(2) Conversely, in liver metastases, blood vessels aggregated at the tumor invasion front. We describe this aggregation of blood vessels as an "excess", because it is statistically significantly higher than expected under the null hypothesis (see Material and Methods section). However, this term does not imply that these blood vessels are "unnecessary"-rather, the biological significance of this pattern must be analyzed in further studies. Some of the possible hypotheses that arise from this pattern are discussed below: For example, the unexpected juxtaposition of the angiogenic belt to the gut lumen could not be explained by ulceration or chronic inflammation in histopathological re-evaluation of the slides.

Possible clinical implications and perspectives
These findings led us to a new model of tumor vascularization (Fig 7A and 7B): a highly vascularized zone of approx. 1.5 mm width is present close to the intestinal lumen in CRC primary tumors. Also, a highly vascularized zone of approx. 1 mm width is present close to the invasion front in CRC liver metastases. This model has several potential implications: 1. Tumor-host-interaction at the intestinal lumen: Several studies have proposed mechanisms of tumor-host interaction of CRC at the basolateral [30][31][32] and on the luminal side [33,34]. Our findings show that CRC tissue facing the intestinal lumen is usually hypervascularized. This raises new questions about interactions at this side of the tumor, e.g. about cancer-microbiome-interactions. It is conceivable that luminal hypervascularization influences diffusion of bacteria-derived or nutrition-derived molecules into the tumor tissue, but vice versa these factors could also contribute to the formation of this new type of spatial vascular heterogeneity in CRC. Based on our findings, it would be highly interesting to investigate molecular interactions between the microbiome, nutrition and the luminal tumor vasculature. Furthermore, it would be interesting to analyze in future studies tumors of hollow organs that face a lumen with a sterile fluid like transitional cell carcinoma of the bladder.
2. Hypoxia: Tumor vascularization determines tumor hypoxia, which triggers a plethora of signaling pathways [35][36][37]. Previously, Righi et al. have shown that tumors with a relative lack of blood vessels at the invasion front are hypoxic in this area and that this is associated with increased tumor cell budding [38]. Our findings significantly extend this study since we demonstrate that in primary CRCs, vascular patterns at the invasion front markedly differ between two groups of tumors: One group has a relative lack of blood vessels at the tumor invasion front (i.e. a "negative excess") and another group of CRC primary tumors that shows a metastasis-like vascularization pattern with a positive blood vessel excess at the invasion front (Fig 5B, Fig 6B).
3. Metastasis: Liver metastasis in CRC is caused by tumor cells that successfully invade the blood stream and subsequently the liver parenchyma [39]. Our experimental results suggest that CRC tumor cells growing close to the intestinal lumen may have a better access to blood vessels than tumor cells in other regions. Further studies should explore how this property of tumors influences the propensity of tumor cells to metastasize to the liver.
4. Distribution of chemotherapy drugs in tumor tissue: Vascular patterns determine how chemotherapeutic drugs are distributed in tumor tissue [40]. Consequently, according to our data it is likely that such drugs reach the luminal side of CRC tumors much easier than the basolateral side. However, the heterogeneous spatial distribution of blood vessels is generally not considered when thinking about the effects of chemotherapy in tumor tissue [41]. Considering quantitative spatial vascularization patterns could help to optimize tumor therapy (e.g. timing of surgery, since the tumor parts of the deep invasion front may be less sensitive to chemotherapy).
5. Anti-angiogenic therapy: Anti-angiogenic therapy has been proved to have beneficial effects in some CRC patients, but these effects are limited [42][43][44]. How spatial patterns of blood vessel distribution might influence the effectivity of this therapy in CRC is currently unknown.
6. Early symptoms of CRC: From a clinical perspective, the presence of an angiogenic belt zone at the intestinal lumen is a conceivable explanation for the common clinical observation that gastrointestinal bleeding is a frequent symptom of colon cancer and that in consequence, tests for occult fecal blood are useful and sensitive.
7. Architecture of CRC in comparison to normal colon mucosa: Normal colon mucosa is a highly structured, well-vascularized tissue. In Fig 8, the normal blood vessel pattern in colon mucosa is shown (described before by Kachlik et al. [45], reproduced in N = 5 WSI in the present study). Our finding that blood vessels in CRC are preferentially located close to the intestinal lumen qualitatively corresponds to the known blood vessel distribution in normal colon mucosa. (Note that it is not possible to quantitatively analyze blood vessel hotspots in normal colon mucosa in the same way as in CRC because normal colon mucosa is much thinner [approx. 700μm] than CRC tissue [several mm to cm] and because its vascular architecture is governed by the structured architecture with crypts).
It is conceivable that CRC tissue inherits the blood vessel architecture from non-neoplastic normal colon mucosa. In further studies, it could be investigated whether this phenotypical resemblance of vascularization pattern also corresponds to functional similarities in proangiogenic signaling.

Limitations, reproducibility and outlook
We describe a novel angiogenic belt zone in CRC with potential clinical implications. Still, the presented results reflect a morphological finding that raises the need for functional investigations. There are three different compartments in a CRC primary tumor that can be clearly distinguished according to their vascularization pattern: a luminal angiogenic belt zone, a tumor bulk and, in a minority of samples, a thin angiogenic belt at the invasion front. Comparative analysis of these three compartments by 3D-histology, mass-spectrometry and (epi-)genetic analysis following micro-dissection will further elucidate the biological role of different spatial patterns in CRC. Furthermore, future studies could evaluate the described spatial patterns in the light of prognostic biomarkers. This could be addressed by combining survival data with morphological measurements.
Our claim that the spatial patterns are universal features of CRC tissue are based on two independent sample cohorts of a total N = 100 tumors. Still, our findings and our thereon based hypothesis should be validated by other researchers: During our experiments, we noted that the vascular belt zone can be identified by thorough visual examination of CD34-immunostained histological slides (see, for example, Fig 4 and S2 Fig)-and in some cases even in H&E stained samples (not shown). Thus, we expect that visual examination (without full quantitative analysis) is sufficient to qualitatively reproduce our findings (still, of course, we provide all our raw data and source code for others to use, see below).
Finally, it is worth mentioning that the current study focuses on the tumor proper. Analogous spatial statistical studies of the peritumoral microenvironment are obviously mandatory in CRC as well.
Supporting information S1 Table. CRC patient characteristics of both cohorts. Sample IDs starting with "C2" are from the second cohort (validation cohort). Four samples from the primary cohort were excluded after staining (these are not shown): Smp042 was an appendix carcinoma, Smp055 was a lymph node metastasis, Smp059 was a duplicate of Smp058 and Smp064 was insufficiently stained and were removed after assignment of sample numbers. T = local tumor stage, N = lymph node stage, M = distant metastasis stage at time of surgery. In total, N = 100 samples were analyzed. (1) Detail from the original WSI (Smp040), scale bar 300 μm, CD34-immunostaining labels blood vessel endothelium. The area is densely packed with blood vessels. (2) In the WSI, three regions are manually defined: intestinal lumen (green), tumor (black), adjacent tissue (orange). Blood vessels in the WSI are automatically detected and plotted as white points. (3) To generate an internal control, detected blood vessels in a ROI are randomly redistributed N = 100 times using a Monte Carlo procedure so that overall MVD remains constant. (4) For each blood vessel, the distance to the intestinal lumen is calculated. This is repeated for each random pattern. The actual distribution of distances is compared to the distribution of random distances, plotted as a histogram. (B) Identification of excess zones: (5) The distance distribution of the random patterns is subtracted from the distance distribution of the observed patterns. The resulting difference is plotted as an excess histogram E(x). The first peak in E(x) reflects an excess of blood vessels at the luminal side. (6) The width of this peak describes the width of the excess zone while the total peak area describes the excess amount. (7) The excess amount is represented as one bar in a waterfall plot. (TIF) S4 Fig. Blood vessel morphology. (A) Angiogenic hotspot probability map in a representative sample (Smp040), ( Ã ) refers to the level of significance. (B) Each blood vessel within the tumor is represented by a filled circle. The bright gray region shows the adjacent tissue and the dark gray region shows the intestinal lumen. As shown above, blood vessels tend to cluster at the intestinal lumen. Additionally, the solidity (defined as area divided by convex area) is shown for each blood vessel (blue = less solid, yellow = solid). Most non-solid vessels are located at the luminal side. This corresponds to the morphological observation that intratumoral blood vessels at the intestinal lumen tend to be more dilated than blood vessels in other tumor regions.