Unsupervised spatial modelling of blood vessel patterns in ... · angiogenesis, the RGP liver...
Transcript of Unsupervised spatial modelling of blood vessel patterns in ... · angiogenesis, the RGP liver...
Koen Marien1,2, Andrew Reynolds3, Kelly Schats1,2, Laure-Anne Teuwen2,4, Pieter-Jan van Dam4, Luc Dirix4, Mark Kockx2, Steven Van Laere4, Peter Vermeulen4
1Laboratory of Physiopharmacology, University of Antwerp; 2HistoGeneX NV; 3Tumour Biology Team, The Institute of Cancer Research (ICR); 4Translation Cancer
Research Unit, GZA Hospitals St. Augustinus
Discussion It is proposed that whilst DGP liver metastases utilise sprouting
angiogenesis, the RGP liver metastases co-opt pre-existing liver
sinusoidal vessels instead. By applying a clustering method to the
blood vessel objects we now have an objective way of confirming
these observations: RGP liver metastases have a vasculature with
a morphology similar to the normal liver sinusoidal system and
without vascular hotspots. Moreover, the proposed method to
quantify vascular hot spots in tissue sections can probably be
applied to detect heterogeneity in sample cohorts of other cancer
types.
Introduction The liver is a well vascularized organ that frequently hosts metastases in patients with colorectal adenocarcinomas (CRC). Different growth patterns at the tumour–liver
interface have been described: desmoplastic (DGP), pushing and replacement (RGP) (Van den Eynden et al., 2013). While the DGP is characterized by desmoplasia,
inflammation and, importantly, sprouting angiogenesis, in the RGP cancer cells “replace” the hepatocytes and co-opt the sinusoidal blood vessels of the liver without eliciting
sprouting angiogenesis. Moreover, our unpublished data suggests that patients with RGP liver metastases respond poorly to bevacizumab, when compared to patients with
DGP liver metastases. This is most likely because bevacizumab can only inhibit sprouting angiogenesis and does not target the co-opted sinusoidal blood vessels. In order to
provide further evidence that the mechanism of tumour vascularisation is different in DGP metastases when compared to RGP metastases, in the current study we
performed unsupervised spatial modelling of blood vessel patterns in patient samples of CRC liver metastases.
Unsupervised spatial modelling of blood vessel patterns in colorectal
cancer liver metastases: additional evidence for non-angiogenic growth
Fig. 1: Overview of the steps done in Definiens to get the coordinates of the segmented vessel objects. Left: Manual region of interest (ROI) delineation by the pathologist in the whole-slide image (WSI) of the liver CRC metastasis. Mid: Threshold-based vessel segmentation with Definiens in the liver (top) and in both the DGP (bottom left) and RGP (bottom right) ROI. Right: Export of the coordinates of the centroids of all vessel objects for post-processing in R (see Fig. 2).
Fig. 3: Unsupervised spatial modelling show similar blood vessel patterns for the RGP of CRC metastases and normal liver. A: Selected ROIs at the tumor-liver interface of normal liver (top), RGP (mid), and DGP (bottom) in CD31-stained tissue. B: Vessel segmentation and classification results in Definiens (Fig. 1). C: Cluster results for the selected ROIs as calculated in R (Fig. 2).
Fig. 4: Normalized number of clusters of blood vessel objects for DGP, RGP and normal liver. The number of clusters was different between DGP and RGP (p < 0.05), but also between DGP and normal liver (p < 0.001). However no difference was found between RGP and normal liver (p = 0.16).
Results There was a statistically significant difference between the growth patterns as determined by one-way
ANOVA (F(2,22) = 10.8, p < 0.001). A post-hoc Tukey test showed that the number of clusters divided
by number of vessel objects (normalization) was significantly different between DGP and RGP (p <
0.05), but also between DGP and normal liver (p < 0.001). However, no difference was found
between RGP and normal liver (p = 0.16).
1. Van den Eynden GG, et al. The multifaceted role of the microenvironment in liver metastasis. Cancer Res. 2013.
2. Lopez XM, et al. Clustering methods applied in the detection of Ki67 hot-spots in whole tumor slide images. Cytom A. 2012.
DGP RGP liver
Nor
mal
ized
nu
mb
er o
f cl
ust
ers
Fig. 2: Overview of the steps done in R to get clustered vessel objects. Creating a cluster is an iterative process. A vessel object is a core-object (red) when it has a least three neighbours in its spherical neighbourhood (black) (radius = maximum nonoutlier value). A density-reachable object (green) is a vessel object that is one of the minimum three neighbours of a core-object, but is not a core-object itself. Together with the core-object, the density-reachable objects belonging to the same neighbourhood define a specific cluster. (Lopez et al., 2012).
Materials and methods
DG
P
RG
P
liver
ns