System Modeling and Machine Learning in Prediction of Metastases

in Lung Cancer

Andrzej Swierniak

, Emilia Kozłowska

, Krzysztof Fujarewicz

, Damian Borys

Agata Wilk

, Jaroslaw Smieja

and Rafal Suwinski

Department of Systems Biology and Engineering, Silesian University of Technology, Akademicka 16,

44-100 Gliwice, Poland

The 2nd Radiotherapy and Chemotherapy Clinic, M. Sklodowska-Curie National Research Institute of Oncology,

Gliwice Branch, Wybrzeze Armii Krajowej 15, Gliwice, Poland

rafal.suwinski@io.gliwice.pl

Keywords: Medical Image Processing, NSCLC, AI Based Models, Metastases, Survival Analysis.

Abstract: The aim of this paper is to present goals and preliminary results of our project devoted to system engineering

approach in prediction of metastases in lung cancer. More specifically we consider existing and develop new

methods of system modeling, machine learning, signal processing and intelligent control to find biomarkers

enabling prediction of risk of tumor spread and colonization of distant organs in non-small-cell lung

carcinoma basing on clinical data and medical images. The results could bring us knowledge about the

dynamics and origin of metastatic dissemination of lung cancer. By dynamics, we understand when and where

a tumor will disseminate, and by origin we mean dissemination path (directly from original tumor or through

lymphatic nodes). This information is very valuable for clinicians, as it could guide the personalized treatment

of lung cancer patients. The results will elucidate important issues concerning prediction of individual

progress of cancer and treatment outcome in oncology. They will provide both theoretical and simulation

tools to support decision making and diagnostics in oncology, on the basis of individual patient state.

1 INTRODUCTION

In this paper we describe main goals and methods

used in a project in which we combine system

engineering methodology with clinical data to predict

metastases in lung cancer. The interest of proposing

original models and methods is to support analysis of

clinical and imaging data and aim at better prediction

of spread and colonization of tumor cells to distant

organs, with emphasis on the most common subtype

of lung cancer - non-small-cell lung carcinoma

(NSCLC). Since the metastatic tumor is mainly

incurable, due to its resistance to treatment, our

research is directed to answer the following urgent

https://orcid.org/0000-0002-5698-5721

https://orcid.org/0000-0002-3069-3085

https://orcid.org/0000-0002-1837-6466

https://orcid.org/0000-0003-0229-2601

https://orcid.org/0000-0001-7554-1803

https://orcid.org/0000-0002-6120-4424

https://orcid.org/0000-0002-3895-7938

biological and clinical question: how, when, and

where the primary tumor will spread to distant

locations. The proposal is focused on reducing

probability of metastasis and evolution of cancer in

distant sites and emergence of evolving resistance to

therapies. To verify applicability of these methods, all

theoretical considerations are related to clinical data

and radiological images, to which we have access. To

reach that goal, both experimental and system

modeling methods are employed in order to meet the

following intermediate objectives:

• Analysis of available radiomic data

incorporated in Positron Emission Tomography/

Computed Tomography (PET/CT) images, and

220

Swierniak, A., Kozłowska, E., Fujarewicz, K., Borys, D., Wilk, A., Smieja, J. and Suwinski, R.

System Modeling and Machine Learning in Prediction of Metastases in Lung Cancer.

DOI: 10.5220/0011705300003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 3: BIOINFORMATICS, pages 220-227

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

application of signal processing and statistical

inference tools to develop original estimators

supporting prognosis of tumor spread to local

lymph nodes and distant organs.

• Development of stochastic compartmental

models based on branching birth-death

processes in which the primary tumors can

metastasize to local lymph nodes and next,

distant metastases can emerge in liver, brain,

and bones. Inter-patient heterogeneity is

accounted by assuming statistical distributions

of model parameters, using the mixed-effects

statistical framework.

• Modification of existing mathematical models

based on ordinary and/or partial differential

equations describing cancer growth and therapy

aimed at taking into account both local and

distant metastases. Estimation of model

parameters is based on available clinical data.

• Development of machine learning tools required

for integration of radiomic and clinical data with

mathematical models mentioned above.

• Modification of models based on evolutionary

game theory supporting analysis of interaction

of different cancer cells phenotypes leading to

emergency of metastasis and resistance to

treatment.

2 BIOMEDICAL BACKGROUND

AND JUSTIFICATION OF

TACKLING SCIENTIFIC

PROBLEM

Lung cancer is one of the most commonly diagnosed

cancer and is the leading cause of cancer-related

deaths. The most common histological subtype is

non-small-cell lung carcinoma (NSCLC), accounting

for 85% of all lung cancer cases (Inamura, 2017).

Advanced NSCLC is more likely to metastasize,

leading to severe symptoms and a decrease in overall

survival. The presence of distant metastases is one of

the most predictive factors of poor prognosis (Popper,

2016). Distant metastases (distant cancer) refer to

cancers that have spread via blood or lymphatic

vessels from the original location (the primary tumor)

to distant organs or lymph nodes. The main cause of

cancer death is associated with metastases, which are

mainly incurable. Thus, distant cancer is resistant to

treatment intervention. Even though cancer

researchers have made a lot of effort to understand the

appearance of metastases, only few preclinical studies

about metastases were translated to clinical practice.

The proposal aims at tackling metastases in the

most common type of lung cancer. If successful, the

project outcome will be information about the

dynamics of tumor metastases in lung cancer, i.e.,

when, where, and how the primary tumor will

metastasize. Information is extracted using a non-

invasive PET/CT imaging techniques. This

information is incorporated in different types of

known and original models using machine learning

tools. The results could bring us knowledge about the

dynamics and origin of metastatic dissemination of

lung cancer. By dynamics, we understand when and

where a tumor will disseminate, and by origin we

mean dissemination path (directly from original

tumor or through lymphatic nodes). This information

is very valuable for clinicians, as it could guide the

personalized treatment of lung cancer patients.

3 MODELING METASTASIS –

METHODS AND TOOLS

Metastasis is a complex process that involves the

spread of a cancer to distant parts of the body from its

original site. In order to become clinically detectable

lesions, it must complete a series of steps at multiple

temporal and spatial scales. The deterministic

description of this process is based on either ODE or

PDE modes. Saidel et al. (Saidel et al, 1976) proposed

a compartmentalized translational ODE model of

metastasis distribution over the time. An important

contribution, used as a basis of many subsequent

works, in the field of modeling metastasis was

introduced by K. Iwata et al. (Iwata et al, 2000). Their

model for the colony size distribution of multiple

metastatic tumors raising from untreated tumor is

represented by the hyperbolic PDE. Model by Iwata

et al. was further analyzed and extended by Barbolosi

et al.(Barbolosi et al, 2009), Devys et al. (Devys et al,

2009), and Benzekry (Benzekry E., 2011). In (Iwata

et al, 2000, Barbolosi et al, 2009, Devys et al, 2009)

the primary tumor is subject to the Gompertz law,

while in (Benzekry E., 2011) the primary tumor is

described by model of tumor growth including

angiogenesis (Hahnfeldt et al, 1999). This model was

developed further by the same group (Benzekry E. et

al, 2016) in conjunction with clinical data and the

mathematical formulation of a metastatic

dissemination. A different, hybrid approach to the

problem of modeling of invasive cancer and

metastases was introduced by Franssen et al.

(Franssen et al, 2019). The authors presented the

general spatial modelling framework of the metastatic

System Modeling and Machine Learning in Prediction of Metastases in Lung Cancer

221

spread of cancer. Their model was then simulated

using clinical data from breast cancer patients and

data of metastatic sites (bones, lung and liver). The

model (Franssen et al, 2019) was further used by a

group of Benzekry in (Bilous et al, 2019, Nicolo et al,

2000). In (Nicolo, 2000) the authors compare

predictions of the metastatic relapse given by a

machine learning and mechanistic modeling

techniques. In (Bilous et al, 2019) a model of the

dynamics of brain metastasis in NSCLCS is

discussed. To our knowledge, this is the only work in

which explicit metastasis in non-small cell lung

cancer is taken into consideration in tumor dynamics

modeling. In (Smieja et al, 2022) we have proposed

probably the simplest model of tumor progression

including metastasis. At the one hand, it contains the

minimum number of compartments and parameters.

At the other hand, it is able to represent

heterogeneous response treatment in a population of

patients and provide a good fit to clinical survival

curves or progression including metastasis which

enables estimation of parameters based on clinical

data.

Stochastic modeling has been used in

mathematical oncology for a relatively short time,

and includes various techniques to take into account

randomness in the process such as tumor progression

or metastasis. Indeed, tumor growth is a random

process as each tumor cell have different cell cycle

length due to internal (process of DNA repair) or

external factors (competition for space and

resources). Thus, stochastic mathematical models

provide a powerful toolbox for mechanistic modeling

of cancer. We have developed a model of NSCLC

progression and dissemination to local lymph nodes

and distant sites (Kozłowska and Swierniak, 2022).

The model is in the form of stochastic

multicompartmental birth-death branching process

model. The branching process is powerful tool in

modeling various processes in biology, especially in

cancer (see e.g. (Kimmel and Axelrod, 2015)). This

mathematical framework is also useful for

mathematical modeling of local metastases, as shown

in (Haeno et al, 2012) by modeling metastases in

pancreatic cancer. The structure of the proposed

model is presented in Figure 1.

The model considers Gompertzian growth of

lung cancer cells from a single cell of Type I, which

does not have the ability to metastasize. The cell,

however, has accumulated all necessary aberrations

needed for proliferation and has fitness advantage

over healthy cells. Type I cell is also treatment-naïve

and thus sensitive to chemo- and radiotherapy. At

each discrete time point (representing the moment of

cell division), the cell can divide or die, and the

population of cells grows according to Gompertzian

growth law. In addition, a cell has small probability

to mutate to Type II cell, which has metastatic

potential, during division. Type II cell is an

aggressive type of cells, thus its growth dynamics is

exponential. This new type of cell appears with

probability u per cell division, as shown in Figure 1

A. Type II lung cancer cell can undergo a process of

dissemination with probability m per cell division,

leading to appearance of a new lung cancer cell in one

of metastasis sites. We assume that the cells in

metastasis sites are resistant to standard treatment,

which is composed of chemotherapy combined with

radiotherapy.

Figure 1: Mathematical model of NSCLC progression and

dissemination. A. Each cell in primary tumor compartment

(lung) can undergo one of three processes division, death,

and dissemination to local/distant site. Each cell in

metastasis compartment can divide or die. We do not

consider secondary metastases. B. Two paths of metastatic

dissemination. Blue arrows indicate dissemination through

lymphatic vessels and red ones through blood vessels.

The model considers two ways of metastatic

dissemination: through blood vessels (hematogenous

route) and through lymphatic vessels (lymphatic

route), as shown in Figure 1 B. The lymphatic route

is shown in blue color, whereas hematogenous route

is depicted with red color. In the first route of

dissemination, lung cancer cells disseminate first

through lymph nodes (local metastases), and next to

one of three distant sites: brain, liver or bones. Those

three distant sites characteristic for NSCLC

metastasis. The hematogenous route of tumor

dissemination is modelled as a single step process

where a Type II lung cancer cell colonizes one of

three distant site brain, liver or bones with probability

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

222

m. Global sensitivity analysis of a preliminary version

of this model was performed in (Kozłowska and

Swierniak, 2022). We found four parameters that

affect MFS: the growth rate of the primary tumor, the

growth rate of distant metastases, dissemination rate

from the primary tumor to distant metastases, and

carrying capacity. From all four parameters, we can

control two of them (to some extent): the growth rate

of the primary tumor (using chemotherapy) and

carrying capacity (using antiangiogenic therapy).

Yet another possibility for modeling invading

and colonization of distant organs is given by

evolutionary games. Evolutionary game theory

(EGT) combines mathematical tools of theory of

games with Darwinian adaptation and species

evolution and may be applied to analysis and

simulation of evolutionary changes within different

subpopulations due to interactions between them. The

result of these interactions (and, possibly, the effect

of environment) is a change of the degree of

evolutionary adjustment which, in turn, may cause

stabilization of the population structure. Using EGT,

it is possible to foresee, whether a population tends to

be heterogeneous or rather only one phenotype

survives and dominates. Introducing changes of the

replicator equations (RE) describing the behavior in

the population in time allows to follow dynamics of

changes. EGT has also been applied to study

development of cellular populations since cells, like

whole organisms, compete for space and nutrients,

exchange signals, cooperate, and show kinds of

“altruism” resembling animals in evolution. Starting

from the pioneering work of Tomlinson and Bodmer

(Tomlinson and Bodmer, 1997) this machinery was

used to model different tumor related phenomena.

Basanta et al. (Basanta et al, 2008) were probably the

first to use this machinery in modeling phenomena

leading to tumor cell invasion and migration. The

authors assume that at initial stage cancer cells are

specified by autonomous growth and then they can

switch to anaerobic glycolysis or become

increasingly motile and invasive. It allows to study

the circumstances, under which mutations confer

increased motility to cells needed for invasion of

other tissues and metastasis. In their next paper

(Basanta et al, 2010), the authors extended their

model by adding phenotype which could switch to

anaerobic glycolysis and be motile. Their model is

directed to glioblastomas. EGT is based on the

assumption of perfect mixing inside the population

(mean field approach) and interaction of each pair of

strategies.

To overcome this simplification and enable

analysis of local arrangement and internal

interactions in the neighborhood, the evolutionary

games have been transferred into spatial lattice by

application of cellular automata techniques, leading

to the so called spatial evolutionary game theory. In

(Swierniak and Krzeslak, 2013) the analysis of all

these three models is appended by RE and SEGT

tools (if absent in original study) which allows to give

an approximate answer on questions regarding time

and place of the switch, leading to tumor migration.

In the project we propose more complex EGT models

of tumor- tumor cells interactions containing different

strategies of dissemination of NSCLC which will take

into account results of other tasks in the project.

Moreover, we apply new tools of spatial evolutionary

tools, proposed recently. These tools take into

account heterogeneity at the cell level (the so called

Mixed Spatial Evolutionary Games – MSEG) and

varying in time (and possibly also in space) effects of

environment (Evolutionary Games with Resources

and Spatial Evolutionary Games with Resources,

respectively). In the former case it leads to multilayer

structure of the game (Swierniak and Krzeslak, 2016)

and in the latter case to time varying pay-off tables

(Swierniak et al, 2018). Moreover, we propose new

algorithms which enable modeling of 3D structure in

spatial games.

4 IMAGE PROCESSING,

FEATURE EXTRACTION AND

SELECTION, MACHINE

LEARNING BASED MODEL

Positron Emission Tomography/Computed

Tomography (PET/CT) examination is currently

routinely used in radiation treatment planning and

staging of patients with NSCLC. It allows for

relatively precise assessment of primary tumor

volume and volume of involved mediastinal lymph

nodes. Retrospective data (including PET/CT images

and history of the treatment) for at least 100 patients

with stage IIIAN2-IIIB NSCLC, who had pre-

treatment PET/CT imaging and underwent curative

radio-chemotherapy have been selected and acquired

from the database. For all those patients all clinical

information will be extracted. Images acquired from

PET/CT device and stored in a standard DICOM

format (Digital Imaging and Communications in

Medicine) will be processed to obtain a set of

radiomics features. Target lesions, for primary tumor,

as well as for nodes or metastatic lesions, are prepared

manually by an experienced specialist, using medical

image viewer software and/or automatically extracted

System Modeling and Machine Learning in Prediction of Metastases in Lung Cancer

223

directly from image, if needed. All regions of interest

(ROI) are stored for subsequent analyses. With high-

throughput computing, it is now possible to rapidly

extract a vast number of quantitative features from

tomographic images (computed tomography (CT),

magnetic resonance (MR), positron emission

tomography (PET)).

The main concept behind this process was that

biomedical images contain information reflecting

underlying pathophysiology and that these

relationships can be revealed via quantitative image

analyses. The conversion of digital medical images

into mineable high-dimensional data is known as

radiomics (d’Amico et al, 2020, van Griethuysen et

al, 2017, Gillies et al, 2016, Kumar et al, 2012,

Lambin et al, 2012). Radiomics is designed to

develop decision support tools; therefore, it involves

combining radiomic data with other patient

characteristics (clinical, molecular etc.), if available,

to increase the power of the decision support models.

It has been proven that features not perceptible to the

eye of the reporting physician — such as intra-tumor

heterogeneity, distribution of signal values within the

tumor area and more — can be indicative of certain

biological characteristics of the tissue, such as

proliferation, hypoxia, necrosis, angiogenesis and

even tumoral genotype (d’Amico et al, 2020).

Quantitative image features based on intensity, shape,

size or volume, and texture offer information on

tumor phenotype and microenvironment that is

distinct from that provided by clinical reports,

laboratory test results, and genomic or proteomic

assays. These features, in conjunction with other

information, can be correlated with clinical treatment

outcomes data and used for clinical decision support

(Figure 2). Radiomics provides imaging biomarkers

that could potentially aid cancer detection, diagnosis,

assessment of prognosis, prediction of response to

treatment, and monitoring of disease status etc.

Acquired pre-treatment PET/CT images are

preprocessed in order to save the data in appropriate

format for subsequent radiomics analysis. Manually

or automatically generated ROIs are preprocessed in

the similar way. Radiomic features are extracted from

the target lesions (described by ROIs) using the

program based on PyRadiomics (https://

pyradiomics.readthedocs.io) package for Python (van

Griethuysen et al, 2017), including: (i) First order

features (energy, entropy, minimum, percentiles,

maximum, mean, median, interquartile range, range,

standard deviation, skewness, kurtosis, among

others); (ii) Shape Features (volume, surface area,

sphericity, among others); (iii) higher order statistics

texture analysis, including: Gray-Level Co-

occurrence Matrix (GLCM), Grey-Level Dependence

Matrix (GLDM), Grey-Level Run Length Matrix

(GLRLM), Grey-Level Size Zone Matrix (GLSZM)

and Neighboring-Gray Tone Difference Matrix

(NGTDM). Using additional filters (for example

Local Binary Patters, wavelets etc.) on the original

image is also considered. This allows to multiply the

resulting radiomic features, which can potentially

highlight features invisible in ROIs.

Figure 2: Illustration of a typical workflow for radiomics

signature development: I group selection, II Data

acquisition, III Image segmentation, IV Feature extraction,

V model validation. Model can be build using not only

radiomic features.

Machine learning algorithms could be applied in two

ways:

a. Classical approach - as a separate algorithm /

predictor whose output is the predicted time and place

of cancer metastasis.

b. Non-classical approach - as part of a larger

model, part of which is one of the dynamic models

developed in the project.

In the first case, the problem is formulated in a

typical way for supervised learning, in which we have

a learning set in the form of PET / CT images,

radiomic features extracted on their basis, and

additional clinical data for a given patient cohort. This

data is accompanied by the set outputs of the model

in the form of information about the time and places

of metastasis. In this situation, the classic division of

the model structure into (i) feature selection and (ii)

training the classifier (predictor) may apply. In

addition, the problem of simultaneous use of radiomic

data and clinical data whose nature is different may

be interesting. In this context, we have tested different

ways of integrating this data and choose the best one.

For this purpose, we use the software, which allows

testing various methods of data integration, while

protecting against information leakage, which may

result in an optimistic bias of prediction quality

assessment. In (Fujarewicz et al, 2022) we have

presented the attempt to use the radiomics features to

predict the metastasis for lung cancer patients. The

I II III IV V

PET/ CT ROI

.csv

ﬁle

model

radiomic s

features

ML / AI

clinical

information

molecular

data

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

224

obtained accuracy of the best classifier confirms the

potential of such prediction of metastasis.

The non-classical approach (b) to the application

of machine learning algorithms and feature selection

relies on the construction of a combined model, part

of which is the dynamic models developed in the

project. In the combined model, a task of the machine

learning algorithm (instead of metastasis prediction)

determines the values of parameters (different for

each patient) of dynamic models. It seems that this

approach, although more difficult than the classical

approach, has a chance of better prediction, because

it combines the advantages of machine learning and

modeling of dynamic systems. The difficulty in

building such a model lies in the fact that the

described approach cannot build a typical supervised

learning data set. While we have input data (radiomic

and clinical), we do not have the set values of these

parameters, but only the set (observed) responses of

patients to therapy. The model learning process must

therefore take account of this fact and requires

development of new learning / adaptation methods.

In the case of the classic approach to the use of

artificial intelligence algorithms (in which the output

of the artificial intelligence algorithm is the predicted

time and place of metastasis) the model learning task

can be formulated as a regression / approximation

task, with continuous output variables, or as a

classification task, in which the model output is the

label of the appropriate class. It is also possible to

create a hybrid model that has both continuous (time

to metastasis) and discrete (place of metastasis)

outputs. In all these cases, we test various regression

and classification techniques such as: support vector

machines (SVMs) with different kernels, linear and

logistic regression, convolution neural networks,

linear (LDA) and quadratic (QDA) discriminant

analysis, classifier ensembles (bagging, boosting,

random forests) and others.

An important element of building the machine

learning model is the selection of features

(radiometric, clinical). In this case, we use various

approaches, ranging from the simplest filter methods

to more complex wrapped methods and embedded

methods. It is also possible to use methods for

transforming feature spaces such as Principal

Components Analysis (PCA) or Independent

Components Analysis (ICA). Nevertheless SVM is

our first choice in extraction and selection of radiomic

features.

In the case of a non-classical approach, in which

the outputs of the machine learning algorithm are the

parameters of the dynamic model, in general case, it

is impossible to use available methods of supervised

learning because (desired) parameters of the dynamic

model are not known and only desired response of the

dynamic model (i.e. patient’s response) is given. In

this case, we present the machine learning task as a

mathematical programming task – optimization of the

performance index depending on the prediction error

in the parameters space of the particular machine

learning method.

In the special case, when the dynamic model has

a form enabling to build based on it the sensitivity

model, we develop an original gradient algorithm

based on the backpropagation of the prediction error

(through the model adjoint to the dynamic model)

enabling the determination of the gradient of the

performance index with respect to the parameters of

the machine learning algorithm. In some respects,

such an algorithm is similar to the algorithms

developed earlier, which involve the use of adjoint

sensitivity analysis for complex systems (Fujarewicz

and Galuszka, 2004, Fujarewicz et al, 2007).

Selection of features, as in the case of the classical

approach, is possible using filter or wrapped methods.

5 CONCLUSIONS

Although the first attempt to use mathematical

modeling to study quantitatively metastases of

untreated lung cancer had more than sixty years of

history (see, (Colins, 1956)), there are currently no

mechanistic models incorporating biomarkers, which

could play a role of prognosis tools that could inform

when and where NSCLC may metastasize. Such tools

could be of great interest to clinicians, supporting

treatment decisions, such as whether to use systemic

therapy or not and with what intensity and duration.

We hope that by extraction of radiomic features from

PET/CT images, their selection and incorporation in

existing and newly built mechanistic models,

predicting NLSCLC spread and metastatic

dissemination will become possible.

On the other hand, there exist many studies in

which empirical models based on statistical data are

used to predict the risk of metastases taking into

account different genomic or proteomic features of

patients. Those studies are related to genotyping and

genomic profiling (e.g. (Li et al, 2013)), expression

of multiple mRNA markers in bronchoscopy (e.g.

(Suwinski et al, 2012)), gene polymorphisms (e.g.

(Butkiewicz et al, 2015)), blood serum proteins (e.g.

(Suwinski et al, 2019)) or more general serum

comparative analysis (e.g. (Pietrowska et al, 2014)),

to mention only a few of them. The approach

proposed in the paper is the first step in construction

System Modeling and Machine Learning in Prediction of Metastases in Lung Cancer

225

of mixed models, which combine mechanistic

models of tumor dynamics with machine learning

models and using data from diagnostic investigations

(in this case biomedical images).

Dynamical models of cancer growth based on

ordinary differential equations (ODE), partial

differential (PDE) and other structured or agent-based

models (see, (Swierniak st al, 2016, Ledzewicz and

Schaettler, 2015, Clairambaut, 2014), for survey),

usually concentrate only at local tumor eradication,

some additionally take into account the surrounding

tissue. Those models do not take into account cancer

reappearance in distant sites after treatment.

Moreover, there is no available mathematical model

taking into account an intermediate step of metastasis

dissemination, which is spread of tumor cells to local

lymph nodes. The mixed machine learning and

mechanistic model proposed by us could be applied

also to other types of solid cancers such as rectal, head

and neck or breast cancer, which also have high

metastasis potential. Thus, in the future, we plan to

extend the method to other types of solid cancers.

In addition, we plan to incorporate molecular

data from liquid biopsy (see e.g. (Suwinski et al,

2019)). Proteomics profiling of blood serum from

about 100 non-small cell lung cancers will be

performed, which will allow to incorporate molecular

features into prediction of distant metastases. In

(Jaksik and Smieja, 2022) we have presented an

attempt to identify which -omics dataset or

combination of them, provide the most relevant

information for the prognosis of lung cancer survival.

This, will enable integration of biomedical images

with molecular data. However, this work is beyond

the scope of this paper.

ACKNOWLEDGEMENTS

This work was supported by Polish National Science

Centre, grant number: UMO-2020/37/B/ST6/01959

and Silesian University of Technology statutory

research funds.

REFERENCES

d’Amico, A., Borys, D., Gorczewska, I. (2020). Radiomics

and artificial Intelligence for PET imaging analysis.

Nuclear Medicine Review, 23, 1: 36–39

Barbolosi, D., Benabdallah, A., Hubert, F., Verga, F.

(2009). Mathematical and numerical analysis for a

model of growing metastatic tumors. Math Biosci

218(1):1-14

Basanta, D., Hatzikirou, H., Simon M, and Deutsch, A.

(2008). Evolutionary game theory elucidates the role of

glycolisys in glioma progression and invasion, Cell

Prolif., 41:980-987.

Basanta, D., Scott, J. G., Rockne R., Swanson, K.R., and

Anderson, A. R. A. (2010) The role of IDH1 mutated

tumor cells in secondary glioblastomas: an evolutionary

game theoretical view, Phys. Biol. 8: 015016.

Benzekry, S. (2011). Mathematical analysis of a two-

dimensional population model of metastatic growth

including angiogenesis. J Evol Equ 11(1):187-213.

Benzekry, S., Tracz, A., Mastri, M., Corbelli, R., Barbolosi,

D., Ebos, J.M. (2016). Modeling spontaneous

metastasis following surgery: an in vivo-in silico

approach. Cancer Res 76(3):535-547.

Bilous, M., Serdjebi, C., Boyer, A., Tomasini, P.,

Pouypoudat, C., Barbolosi, D., Barlesi, F., Chomy, F.,

Benzekry, S. (2019). Quantitative mathematical

modeling of clinical brain metastasis dynamics in non-

small cell lung cancer. Sci. Rep. 9:13018.

Butkiewicz, D., Krzesniak, M., Drosik, A., Giglok, M.,

Gdowicz-Kłosok, A., Kosarewicz, A., Rusin, M.,

Masłyk, B., Gawkowska-Suwinska, M., Suwinski, R.

(2015). The VEGFR2, COX-2 and MMP-2

polymorphisms are associated with clinical outcome of

patients with inoperable non-small cell lung cancer, Int.

J.Cancer, 137: 2332–2342.

Clairambault, J. (2014). Deterministic mathematical

modelling for cancer chronotherapeutics: cell

population dynamics and treatment optimisation. In

Mathematical Oncology 2013, A. d'Onofrio, A.

Gandolfi Eds., Part III:265-294, Birkhäuser, New York.

Collins, V.P., Loeffley, K.R., and Tivey, H (1956).

Observations on growth rates of human tumors, Am. J.

Roentgen, 76, 988-1002

Devys, A., Goudon, T., Lafitte, P. (2009). A model

describing the growth and the size distribution of

multiple metastatic tumors. Discrete Cont Dyn-B

12:731-767.

Franssen, L. C., Lorenzi, T., Burgess, A. E. F. & Chaplain,

M. A. J. (2019). A Mathematical Framework for

Modelling the Metastatic Spread of Cancer. Bull. Math.

Biol. 81:1965

Fujarewicz, K., Galuszka, A. (2004). Generalized

backpropagation through time for continuous time

neural networks and discrete time measurements., In:

Lecture Notes in Computer Science (eds: L. Rutkowski,

J. Siekmann, R. Tadeusiewicz and L. A. Zadeh), 3070:

190-196, Springer. 2004.

Fujarewicz, K., Kimmel, M., Lipniacki, T., Swierniak, A.

(2007). Adjoint systems for models of cell signaling

pathways and their application to parameter fitting.

IEEE-ACM Transactions on Computational Biology

and Bioinformatics, 4(3): 322-335.

Fujarewicz, K., Wilk, A., Borys, D., d’Amico, A.,

Suwiński, R., Świerniak, A. (2022). Machine Learning

Approach to Predict Metastasis in Lung Cancer Based

on Radiomic Features. In: Intelligent Information and

Database

Systems, Nguyen, N.T. et al. (eds): 40-50,

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

226

Lecture Notes in Computer Science, ACIIDS 2022,

Springer, Cham.

Gillies, R.J., Kinahan, P.E. and Hricak, H. (2016).

Radiomics: images are more than pictures, they are

data. Radiology, 278(2),563-577.

van Griethuysen, J. J. M., Fedorov, A., Parmar, C., Hosny,

A., Aucoin, N., Narayan, V., Beets-Tan, R. G. H.,

Fillon-Robin, J. C., Pieper, S., Aerts, H. J. W. L. (2017).

Computational Radiomics System to Decode the

Radiographic Phenotype. Cancer Research, 77(21),

e104–e107.

Haeno, H. et al. (2012). Computational modeling of

pancreatic cancer reveals kinetics of metastasis

suggesting optimum treatment strategies. Cell 148,

362–375.

Hahnfeldt, P., Panigraphy, D., Folkman, J., and Hlatky, L.

(1999). Tumor development under angiogenic

signaling: a dynamical theory of tumor growth,

treatment, response and postvascular dormancy.

Cancer Research, 59:4770–4775.

Inamura, K. (2017). Lung cancer: understanding its

molecular pathology and the 2015 WHO classification.

Front Oncol. Aug 28; 7.

Iwata, K., Kawasaki, K., Shigesada, N. (2000). A

dynamical model for the growth and size distribution of

multiple metastatic tumors. J Theor Biol 203(2):177-

186.

Jaksik, R., Smieja, J. (2022). Prediction of lung cancer

survival basing on -omic data, In: Intelligent

Information and Database Systems, Nguyen, N.T. et al.

(eds):116-127. Lecture Notes in Computer Science, vol

13758. ACIIDS 2022, Springer, Cham.

Kimmel, M., Axelrod, D.E. (2015). Branching Processes in

Biology, Springer, New York, Heidelberg Dordrecht

London.

Kozłowska, E., Świerniak, A. (2022). The Stochastic

Mathematical Model Predicts Angio-Therapy Could

Delay the Emergence of Metastases in Lung Cancer. In:

Biocybernetics and Biomedical Engineering – Current

Trends and Challenges, D.G., Zieliński K., Liebert A.,

Kacprzyk J. (eds). Lecture Notes in Networks and

Systems, vol 293. Springer, Cham.

Kumar, V., Gu, Y., Basu, S., Berglund, A., Eschrich, S.A.,

Schabath, M.B., Forster, K., Aerts, H.J., Dekker, A.,

Fenstermacher, D. and Goldgof, D.B. (2012).

Radiomics: the process and the challenges. Magnetic

resonance imaging, 30(9).1234-1248.

Lambin, P., Rios-Velazquez, E., Leijenaar, R., Carvalho,

S., Van Stiphout, R.G., Granton, P., Zegers, C.M.,

Gillies, R., Boellard, R., Dekker, A. and Aerts, H.J.

(2012). Radiomics: extracting more information from

medical images using advanced feature

analysis. European Journal of Cancer,48(4),.441-446.

Ledzewicz, U., Schaettler, H. (2015). Optimal Control for

Mathematical Models of Cancer Therapies, Springer,

New York.

Li, T., Kung, H.J., Mack, P.H., Gandara, D.R. (2013).

Genotyping and Genomic Profiling of Non–Small-Cell

Lung Cancer: Implications for Current and Future

Therapies, Journal of Clinical Oncology, 31 (8), 1039-

1043.

Nicolo, C., Perier, C., Prague, M., Bellera, C., MacGrogan,

G., Saut, O., Benzekry, S. (2000) Machine Learning

and Mechanistic Modeling for Prediction of Metastatic

Relapse in Early-Stage Breast Cancer. JCO Clin

Cancer Inform, 4:259-274.

Pietrowska, M., Jelonek, K., Michalak, M., Roś, M.,

Rodziewicz, P., Chmielewska, K., Polański, K.,

Polańska, J., Gdowicz- Kłosok, A., Giglok, M.,

Suwiński, R., Tarnawski, R., Dziadziuszko, R.,

Rzyman, R., and Widłak, P. (2014). Identification of

serum proteome components associated with

progression of non-small cell lung cancer, Acta

Biochem. Pol. 61 (2), 325–331.

Popper, H.H. (2016). Progression and metastasis of lung

cancer. Cancer Metastasis Rev. Mar1;35(1):75–91.

Saidel, G.M., Liotta, L.A., Kleinerman, J. (1976). System

dynamics of a metastatic process from an implanted

tumor. J Theor Biol 56(2):417-434.

Smieja, J., Psiuk-Maksymowicz, K., Swierniak, A. (2022).

A Framework for Modeling and Efficacy Evaluation of

Treatment of Cancer with Metastasis. In:

Biocybernetics and Biomedical Engineering – Current

Trends and Challenges. Pijanowska, D.G., Zieliński,

K., Liebert, A., Kacprzyk, J. (eds). Lecture Notes in

Networks and Systems, vol 293. Springer, Cham.

Suwinski, R., Klusek, A., Tyszkiewicz, T., Kowalska, M.,

Szczesniak-Klusek, B., Gawkowska-Suwinska, M.,

Tukiendorf, A., Kozielski, J., Jarzab, M. (2012). Gene

Expression from Bronchoscopy Obtained Tumour

Samples as a Predictor of Outcome in Advanced

Inoperable Lung Cancer, Plos One 7 ( 7): e41379.

Suwinski, R., Giglok, M., Galwas-Kliber, K., Idasiak, A.,

Jochymek, B., Deja, R., Maslyk, B., Mrochem-Kwar-

ciak, J., Butkiewicz, D. (2019). Blood serum proteins as

biomarkers for prediction of survival, locoregional

control and distant metastasis rate in radiotherapy and

radio-chemotherapy for non-small cell lung cancer,

BMC Cancer, 19:427.

Swierniak A., Krześlak M. (2013). Application of

evolutionary games to modeling carcinogenesis, Math

Biosci Eng, 10(3), 873-911.

Swierniak A, Kimmel M, Smieja J, Puszynski K, Psiuk-

Maksymowicz K. (2016). System Engineering

Approach to Planing Anticancer Therapies, Springer,

New York, Heidelberg Dordrecht London.

Swierniak, A., Krzeslak, M., Borys, D., and Kimmel M.

(2018). The role of interventions in the cancer

evolution-an evolutionary games approach. Math

Biosci. Eng.16(10); 265-291.

Swierniak, A., Krzeslak, M. (2016). Cancer heterogeneity

and multilayer spatial evolutionary games. Biology

Direct, 11(1):53-61.

Tomlinson I.P.M., Bodmer W.F. (1997). Modeling the

consequences of interactions between tumour cells

British Journal of Cancer, 75, 1997, 157-180.

System Modeling and Machine Learning in Prediction of Metastases in Lung Cancer

227