Quantification of tumoral tissue vascularization has become important to scientific research since Folkman’s revolutionary idea that no tumoral tissue can grow more than 2 mm without vascularisation [
Nowadays computer analysis of endothelial area became more frequently used in tumoral angiogenesis studies. However, digitization instrumentation, analysis input, their processing workflow, and performed measurements vary from one study to another, and none was yet adopted as a standard procedure for vasculature assessment. Moreover, to our best knowledge, none of them investigated intercompartment comparisons of vasculature parameters [
Computer analysis of virtual slides was used in several studies [
Our aim is to assess vasculature in distinct tumoral compartments in an objective manner using a reproducible automated method. The various tumoral compartments shall be manually marked up on virtual slides by a human expert as follows: invasive front, tumour area, and tumour associated stroma.
We analyzed 50 samples of carcinoma of patients between 37 and 70 years old (mean age 57), diagnosed with breast invasive carcinoma, NST (invasive ductal carcinoma, NOS). The women were without any hormonal or chemotherapy before the surgical resection. For each patient, we gathered additional information like medical pathological records; morphological description, TNM classification, histological grade, ER, PR, HER2/neu expression, molecular subtypes, and various correlations were investigated statistically.
Paraffin embedded tumor blocks were cut (5
Samples were digitized and analyzed with TissueFAXS 3.5 (TissueGnostics Gmbh, Vienna, Austria) which included both the scanner as well as the cytometry analysis packages (TissueQuest and HistoQuest). The system consisted of a Zeiss Imager Z2 microscope equipped with a 3 Megapixel area scan colour camera Pixelink PL-623 CF. The motorized stage from Maerzhauser had an 8-slide insert easing the batch scan. The white light was delivered by a VISLED lamp based on LED technology which ensured a stable reproducible intensity over the entire study. The computer used for analysis was an HP Z400 running on an Intel Xeon W3565 processor at 3.2 GHz under Windows 7 32 bit.
The image acquisition phase was done with a 10x magnification objective. Proper microscope settings were checked every day following Koehler illumination procedures described in Zeiss Imager manuals. The camera sensor was aligned to the stage so that the angle between their axis was less than 0.01 degrees, thus avoiding systematic image alignment errors. TissueFAXS 3.5 scanning software was set to store the image tiles in JPEG format with a 95% quality index. The virtual slides were realized by enabling an image overlap of 15% while the integrated algorithm realigned the fields of view by using the overlapping content. This fine-tuning step corrected minor stage errors (up to 2
An initial contextual user analysis phase included visual assessment of the virtual slides from a pathological point of view. Thus, the sections were investigated for locations of tumour area (parenchyma), tumour associated stroma (TAS), and invasive front. The tumour area was identified having an irregular stellated outline pattern, including epithelial tumour cells describing ducts, nests, and cords. The invasive front was perceived as the interface between the periphery of tumour and the adjacent breast tissue. We observed that the breast tumour growth pattern is characterized most frequently by infiltrating and widespread dissection of normal tissue with loss of clear boundary between tumour and host tissues. In addition, we sometimes noted at high magnification the particular aspect of invasive front, with discontinuous small aggregates or single, isolated tumour cells, pattern also known as “tumour budding” [
For each one of these domains, 1–3 sites were selected and marked for analysis using standard regions of interest (ROI) tool by adding predefined circular shaped ROIs of 1 mm diameter. Each of these ROIs had contours highlighted in green, blue, or red, as they belonged to stroma, tumour, or invasive front, respectively. Setting of such coloured highlighting decreased user errors and improved the time spent on visual assessment and secondary opinion analysis during reaching interexpert agreement as well as during postanalysis checks. ROIs selections were done so that difficult image areas were avoided. Folded tissue or with mechanical disrupted morphology generated by cutting, air bubbles within mounting medium or major staining artefacts were disregarded from the analysis.
Following these rules, two-phase ROI definition was performed. During the first phase, 2 pathologists (CC, MD) independently selected the ROIs. A second review phase included multiple consensual meetings in order to confirm, comment, or change ROI sites upon common agreement in all samples (Figures
Selecting an ROI of tumour associated stroma (TAS) immunomarked with CD34-green circle, having at the same time an overview of the whole sample, and zoom in of ROI.
ROI of tumour (T) area stained with CD34, overview of whole sample, and zoom in target area.
Invasive front (IF) selection area stained with CD34, overview of whole sample, and zoom in marked area.
The computer analysis was done using HistoQuest 3.5 cytometry software. TissueFAXS virtual slides were imported into HistoQuest analysis projects, with reusing shading corrections and image tile overlapping information. The software splits the color of RGB image into marker-specific channels using an integrated colour separation method named single reference shade. This approach can separate Hematoxylin and CD34-DAB stains into their optical density (OD) counterparts after a training procedure which involves pointing with the mouse pixels for each of the two stains. The system does not require preparation of separate samples stained only with one of the two markers as it can compensate also for mixtures of stains used as training data. The colour separation method computes an abundance map for each marker, extracting for each pixel the amount of Hematoxylin and CD34-DAB, respectively. This approach allows the assessment of CD34-DAB pixel intensities independently of other existing counterstain. Having the CD34-DAB abundance maps, simple thresholding can be applied to extract positive areas. Although the software allows automatic controlled thresholding, the preferred method included setting a manual threshold of the CD34-DAB OD for all samples and quantifying endothelial area (EA) using total area measurement option. A manual iterative search of the proper threshold was performed by looking at the image results showing overlays of contours on top of original images of several samples. The interactive threshold selection features of HistoQuest, as well as the possibility to apply the settings on small test regions, allowed for a visual confirmation of selected parameters during the iterative search. A HistoQuest marker profile was created to save all colorimetric and thresholding parameters and was used in all analysis projects of the study.
The analysis of all projects containing both definitions of the ROIs within the domains and the analysis parameters was performed using the batch analysis module of HistoQuest. This allowed automatic unsupervised quantification of the entire data set, which took about 4 hours running on all the available 4 cores of the processor.
Having all projects analyzed, a visual validation of proper EA identification was performed. Each project was reopened, and overlays of vessel contours on the colour images were assessed. Blood vessels which were not identified by the system due to too weak stain intensities were manually indicated using Manual Correction Tool -> Add Event. On the other side, as CD34 is known to be expressed also in fibroblast [
Measurement results were exported from the validated analyzed projects using batch export module which generated a single excel sheet containing all relevant patient data: total analyzed area (AA) and total endothelial area (EA) of each domain of each patient. Derived measurements were computed directly in Excel, one of them being the relative endothelial area (REA) for each domain (REA percent = 100 * EA/AA). Statistical analysis was performed in SPSS 19 using the data from exported excels.
Stromal average vasculature percentage (TAS-REA) was found at 0.91% when taking into consideration all patients. In tumour area the average value was 1.95% (T-REA) and in the invasive front was 4.2% of endothelial area (IF-REA). We noticed an increase of 2.15 times of relative vasculature area in the invasive front when compared to the tumour center. This observed trend confirmed our supposition that domain specific measurement may reveal more localized information about tumour angiogenesis development than whole slide parameters (see Figure
REA values for each compartment (TAS, T, and IF) for all patients, lymph node negative subgroup (N0), and node positive subgroup (N > N0).
REA per compartment (relative endothelial area) | All patients, |
Node negative breast cancer patients (N0 group), 13 cases | Node positive breast cancer patients (N > N0 group), 37 cases |
---|---|---|---|
TAS% | 0.91% | 0.91% | 0.92% |
T% | 1.95% | 2.72% | 1.67% |
IF% | 4.2% | 4.99% | 3.92% |
Average REA values for each of TAS, tumour parenchyma, and invasive front compartments of all patients.
Furthermore, similar statistical analysis was performed for each subgroup of lymph node negative (N0) patients as well as lymph node positive patients (N > N0). Same trends were observed for the three compartments in both groups: 0.91 in TAS, 2.72 in T, and 4.99 in IF for the N0 group and 0.92, 1.67, and 3.92 for the N > N0 group, respectively. This shows an increase of vasculature in the invasive front when compared to tumour or stroma values. However, we noticed that this trend is more pronounced in the N0 group of patients than in N > N0 group (4.99 versus 3.92, resp.) (see Figure
TAS-REA, T-REA, and IF-REA for both N0 and N > N0 group of patients. N0 group developed slightly more vasculature in tumour and invasive front compartments when comparing with N > N0 group.
Statistical comparison of tumour compartments endothelial areas (TAS-REA, T-REA, and IF-REA) determined the following Pearson correlations and statistical differences. We found a positive weak statistical correlation but significant between TAS-REA and T-REA (
Paired samples correlations between tumour compartments in our study lot.
Pair of compartments | No. of cases | Pearson correlation |
Statistic difference Sig |
|
---|---|---|---|---|
|
Statistical significance |
|||
TAS-T | 50 | 0.418 | 0.003 |
0.002 |
TAS-IF | 50 | 0.432 | 0.002 |
0.000 |
T-IF | 50 | 0.655 | 0.000 |
0.000 |
Two tailed
When performing
Two-tailed
Statistical correlations in different tumour compartments (TAS, T, and IF) within breast cancer molecular subtypes.
Molecular subtype of breast cancer | No. of cases | Pair of compartments | Correlation ( |
Statistical significance |
---|---|---|---|---|
Luminal A | 25 |
|
|
|
TAS-IF | 0.284 |
| ||
|
|
|
||
| ||||
Luminal B | 9 | TAS-T | 0.315 |
|
TAS-IF | 0.602 |
| ||
|
|
|
||
| ||||
Basal-like | 5 |
|
|
|
TAS-IF | 0.425 |
| ||
T-IF | 0.640 |
| ||
| ||||
HER2 | 11 | TAS-T | 0.373 |
|
TAS-IF | 0.263 |
| ||
|
|
|
Tumour angiogenesis became a point of interest for medical researchers since 1970 when Judah Folkman formulated the axiom that no tumour (new) tissue can grow more than 1-2 mm without development of new vasculature [
As MVD and Chalkley techniques were extensively used for angiogenesis assessment soon after their introduction, newer studies improved the observer independence by enhancing the methods with various computer-aided image analysis systems (CIAS). The methods build further on the two phases of the analysis: selection of hot spots and assessment of vasculature. Thus, various methods of hotspot selection used lower magnification objectives (i.e., 4x or 5x) and an image processing step which detected locations with higher densities of the endothelial marker. However, they were placed regardless of the tumour domain (tumour associated stroma, tumour center, or its invasive front). Some [
Steps in evolution of vasculature assessment.
Study | Region measured | Parameter recommended | Measured by |
---|---|---|---|
Weidner et al. 1991 [ |
Hot spot at tumor border | MVD | human expert |
Barbareschi et al. 1995 [ |
Hot spots identified on low magnification | EA, MVD | CIAS |
Belien et al. 1999 [ |
Whole slide + hotspot | MVD | CIAS |
Oh et al. 2001 [ |
Random spots | MVD | CIAS |
Kim et al. 2003 [ |
Whole slide + hotspot | MVD + EA | CIAS |
Chantrain et al. 2003 [ |
Entire sample | EA | CIAS |
Mikalsen et al. 2011 [ |
Hot spots identified on low magnification | MVD | CIAS |
Our method | Whole slide + domain specific large hot spots | EA | CIAS + human expert |
As digital image processing algorithms will evolve, our approach could be further improved. The first phase of domain specific region delineation could be performed automatically by dedicated algorithms, thus improving the observer independence. Regarding the measurement phase, further more advanced morphological parameters could be investigated, that is, using fractal and syntactic structure analysis [
A major factor in angiogenesis assessment is represented by the selection of the vessel marker [
Since tumour angiogenesis is enhanced by chemical stimulations, microenvironment may play a critical role in development of vasculature. However, the microenvironment in various morphopathological domains (i.e., tumour associated stroma, tumour area, and invasive front) is known to have high variations of protein density patterns [
Digitization standardization was performed to achieve consistent image quality over the entire study. Virtual slides were realized with whole tissue sections so that any domain selection process is done having the entire imagistic morphological context available. Scanning only several sites selected directly on the microscope by one expert would have deprived the additional human experts from valuable unretrievable image data and would have dramatically restricted the secondary opinion analysis. In our case, the compartment selection phase was fully traceable and changeable during the reviewing process using virtual slide digital annotations.
Our measurements aimed at cell-related parameters instead of targeting vasculature morphological entities typical to MVD count approaches. Therefore, we chose to measure REA of endothelial tissue in various morphopathological sites. This REA successfully characterizes total vasculature as well as provides a reliable base for endothelial proliferation index assessment considering that the area is statistically proportional with cell count. A comprehensive user validation phase was performed by looking at contours of found objects overlayed on the original images. Other image-type results (masks, gray level representations, and so forth) available in the software were also used. This approval phase took into consideration the expected morphology of the vasculature structures and allowed manual removal of falsely found areas which did not resemble blood vessels (i.e., high background, folded tissue, and so forth). Worth to mention that CD34 is known to stain also fibroblasts, which were also manually removed from analysis when found (
After processing the data from quantification of endothelial area of interest, we noticed a significant difference of CD34 in tumour (T-REA = 1.95%) versus stroma (S-REA = 0.91%). The tumour has a vascularisation index 2.14 times higher than the tumour associated stroma index (
Various hot spots based measurements typically assessed with MVD may reveal local densities that are much higher than normal tissue, thus illustrating its malformed organization and highly heterogeneous network [
When analysing the vascularisation according to molecular type of the tumours, we observed that in luminal A type carcinomas the statistical correlations for both pairs of compartments (T-IF and TAS-T) had moderate significance. This suggests that it might be a correlation between the better prognosis of the luminal A type breast carcinoma and the vascularization. On the other hand, for the rest of the molecular types (luminal B, basal-like, and HER2) we found highly significant positive correlations between some pairs of the three compartments. The correlations of different compartment pairs were not consistent between all molecular subtypes. This indicates different vasculature dynamism in each case since distinct variables and mechanisms may be involved. Correlations with other additional molecular signatures should be investigated in larger studies for a better understanding of the found differences.
The variations of REA index found between different tumour compartments, as well as between patient subgroups (lymph node negative versus node positive groups), show that domain specific REA measurements are capable of confirming and revealing additional important information about cancer development. Moreover, it may be an important criterion for further subgrouping and classification within already widely accepted histological scores. Thus, once new targets for cancer treatment are discovered, the proposed method can be used for assessment of the patient outcome.
Relevant improvements in traceability and observer independence were realized by digitization of whole slide and virtual annotation of the domains of interest. The proposed measurement of relative endothelial area index for each of the tumour compartments (tumour associated stroma, tumour parenchyma, and invasive front) showed relevant differences in microvessel local density. It also showed differences between patients with or without lymph nodes metastases. The new digital scoring procedure can provide a precise measurement tool that promotes marker identification and correlation with a significant impact for patient management and eventually treatment individualization. By combining the experience of the pathologist in hot-spot selection with the precise measurement of the image processing approach, the proposed methodology brings new insights in clinical diagnostic, patient treatment, and follow-up evaluation.
The entire current study, TFAXS acquisition, and analysis platform were financed by the Grant CNCSIS code 29 “Interdisciplinary platform of Molecular Medicine” of “Gr. T. Popa” Medical University Iasi. The work is part of PhD thesis of Dr. Anca Haisan, as well as of the PhD thesis of Radu Rogojanu. The authors Anca Haisan, Camelia Croitoru, Daniela Jitaru, Cristina Tarniceriu, Mihai Danciu and Eugen Carasevici declare no financial conflict of interests. Radu Rogojanu, PhD student at Medical University of Vienna working also for TissueGnostics, provided scientific and technical expertise in optimal usage of TissueGnostics product and declared that no financial support was provided to the current scientific work or any of its coauthors.
Anca Haisan and Radu Rogojanu contributed equally to this work.