Digital pathology becomes more and more important nowadays. The field is growing,
as continuous development of modern digital scanner allows the scanning of whole
slides in high resolution. The whole slide images enable the application of
computer-aided methods to support pathologists with additional quantitative information. This
work aims to develop methods for the analysis of whole slide images. 35 tissue sections
of classical Hodgkin lymphoma were analyzed regarding CD30 cell positions and their
morphology using graph-theoretical methods.
Lymph nodes are small, bean-shaped structures and a major part of the lymphatic
system of mammals. The human body contains about 500-700 lymph nodes of different
size. Foreign particles are transported via the lymph fluid to be filtered and recognized
efficiently in the lymph nodes. This enables a quick immune response in case of an
infection.
Hodgkin lymphoma (HL), also called Morbus Hodgkin, is a malignant tumor of the
lymphatic system. In general, the malignant cells in HL derive from B-cell pre-cursor
cells of the germinal center. In some exceptional cases, they have been found to originate
from T-Cells.
A little more than 9,000 new cases are diagnosed annually as Hodgkin lymphoma in
the United States of America. Despite the high 5-years survival rate of 85.3 %, HL is
still a severe disease and causes the death of about 1,100 people per year in the USA.
Additionally, the treatment with chemo- and radiotherapy can cause subsequent diseases.
According the World Health Organisation (WHO), HL can be differentiated in subgroups.
The most frequent one is the classical Hodgkin lymphoma (cHL). Cases diagnosed
as cHL can be further characterized. The two most common types of cHL are nodular
sclerosis cHL and mixed cellularity cHL.
Histopathologically, Hodgkin lymphoma is characterized by the occurrence of Hodgkin
and Reed-Sternberg (HRS) cells. In microscopic images, these cells appear enlarged, and
have one or more large cell nuclei with a coarse-grained chromatin structure. HRS cells
are known to be positive for the activation marker CD30, and thus can be highlighted by
immunostaining. Immunohistological staining is a commonly used method in pathology
and is, in combination with microscopy, routinely used for diagnosis of lymphoma.
Beside traditional microscopy, digital scanners nowadays allow the digitization of
whole object slides to whole slide images (WSIs). The storage efficiency and the usability
in telepathology are the most distinguished advantages of WSIs. Nevertheless, WSIs are
currently not used in routine pathology. The effort and costs to change the established
routine processes in pathology and to integrate digital object slides are still high.
Computer-aided analysis of tissue sections may lead to additional, more sophisticated,
information to describe the diseases’ progress. Statistical data gained from the whole
tissue section can support pathologists to characterize disease patterns in more detail
and may assist the diagnosis.
The aim of this work is the exemplary analysis of WSIs of tissue sections diagnosed as
cHL. The immunohistological images were provided by the Dr. Senckenberg Institute of
Pathology at the Goethe-University Frankfurt. The examined images are immunostained
against CD30. CD30 is a membrane receptor, which is expressed by HRS cells, and
therefore labels the malignant cells in cHL. A second staining, hematoxylin, is applied
to highlight the nuclei of all cells in the tissue section. The WSIs were captured with an
Aperio ScanScope slide scanner at a high resolution of 0.25 μm per pixel. The size of the
tissue sections resulted in images with up to 90,000 x 90,000 pixels. Without compression
the files can reach sizes of 30 GB.
The pre-selected image set consists of 35 WSIs with tissue sections diagnosed as
mixed cellularity cHL, nodular sclerosis cHL and lymphadenitis. The latter is examined
as a control group. Here, the CD30-positive cells are not tumor cells, but activated
lymphocytes. They are part of an immune response against a viral or bacterial infection.
To deal with the non-standard SVS file format of Aperio and for the analysis of the
images, we implemented an in-house software called Impro. The software is written in
Java. The required image processing methods were implemented from scratch. Examples
are the detection of a region of interest (ROI) and the color deconvolution to separate
the stains. Additional methods, like the thresholding for the image segmentation and the
computation of morphological cell descriptors, were integrated using established imaging
software and libraries like CellProfiler and the Java Advanced Imaging API (JAI).
The cell detection pipeline identified more than 400,000 cells in the 35 WSIs. The
number of CD30-positive cells varied among the 35 cases. Overall, the cell count is lowest
in lymphadenitis cases. While lymphadenitis tissue sections had on average 3,000 CD30-
positive cells, in mixed cellularity cHL tissue sections, the average count was 19,000.
A few cHL cases existed for which the number of HRS cells exceeded 50,000. The cell
density can be displayed in Impro as a heat map overlay on top of the tissue section.
In lymphadenitis, the CD30-positive cells were distributed evenly throughout the tissue
section. In contrast, the cells formed dense groups in cHL cases. Especially in nodular
sclerosis cHL, the CD30-positive cells formed dense clusters, even if the total number of
cells in the tissue section was low.
For each CD30-positive cell, the imaging pipeline computed morphological descriptors
like eccentricity, solidity and area size. The descriptors allow a more detailed view of
the disease pattern. Up to now, form and size of HRS cells have been only described
on single, manually selected HRS cells. The presented approach allows a statistical view
on the cell shape. It is noteworthy that the cell detection pipeline does not differentiate
between HRS cells and CD30-positive cells in general. The Feret diameter describes the
extent of an object. CD30-positive cells in cHL cases have an average Feret diameter of
20 μm. The CD30-positive cell population in lymphadenitis has a diameter of 15 μm on
average.
Beside the statistical analysis of single cells, the aim of this work is to model the lymphoma
as a complex system. We applied system biology methods to depict the relations
of neighboring cells. The cell positions are used to build up cell graphs. Neighborhood
relations are modeled according to the unit disk graph formalism. Typical graph properties
can be computed to characterize the different tissue sections. The cHL cases show
an increased average vertex degree compared to lymphadenitis, meaning that the microenvironment
consists of more CD30-positive cells. In mixed cellularity cHL, we also
see a high variability for the vertex degree distributions. Compared to random geometric
graphs, the analyzed cell graphs have an increased average vertex degree. Even in
lymphadenitis, where the CD30-positive cells are more evenly distributed than in cHL
cases, the average degree is higher than one would expect from randomly distributed
cells. Lymphadenitis is a controlled immune response and may be limited to the parts
of the lymph node where the foreign particles were recognized.
Many graphs exhibit a hierarchical structure, in which highly connected vertex groups
exist. In graph theory, these groups with vertices sharing a high number of relations,
are called communities. Clique-based algorithms recognize communities in cell graphs.
The partitioning in cell groups can be displayed as overlay. The number and the size of
communities can be used to characterize cHL tissue sections.
The presented results illustrate that the analysis of WSIs and the additional information
gained from the image processing pipeline can be used to support the diagnosis
of lymphoma. 35 WSIs in total were examined, and a cell detection was performed on
the image layer with highest resolution. More than 400,000 CD30-positive cell objects
were morphologically described. In combination with their position in the tissue section,
important features of disease patterns, e.g. the distribution of malignant cells, becomes
possible. Nevertheless, the proposed imaging pipeline is more or less a prototype. The
application in routine pathology requires a response in a few minutes to be efficient.
Currently, the proposed cell graphs only consist of CD30-positive cells. The method
can be extended in the future and other cell types, e.g. cells that are part of an immune
response, can be integrated. This will allow to analyze the interaction of the tumor cells
and their microenvironment. The cell graph approach can be generalized and allows the
modeling of other disease patterns.