Towards Lung Cancer Staging via Μultipositional Radiomics and

Machine Learning

Dimitris Fotopoulos

, Dimitris Filos

, Ekaterini Xinou

and Ioanna Chouvarda

School of Medicine, Aristotle University of Thessaloniki, Thessaloniki, Greece

Theagenio Cancer Hospital, Thessaloniki, Greece

Keywords: Lung Cancer, Disease Staging, CT Imaging Radiomics, Tumour and Organ Features, Machine Learning.

Abstract: This work addresses lung cancer diagnosis, and more specifically disease staging, as a major clinical challenge,

crucial for further treatment decisions. The procedure is currently performed by experts based on clinical and

medical imaging data and is time consuming and costly. Within INCISIVE, an EU-funded research project

which aims to develop a pan-European federated image repository for cancer and implement Artificial

Intelligence (AI) tools for clinical practice, clinical challenges have been identified that can be supported by

AI in medical imaging data to facilitate accurate diagnosis and support treatment planning. The support and

automation of lung cancer staging has been identified as a priority among the INCISIVE clinical challenges.

In this scope, we propose a method to automatically classify between the group that represents disease stages

I and II (low severity), vs the group that includes stages III and IV (severe). Tumour-Node-Metastasis system

is used as a reference for staging. Based on lung CT image series with tumour and lung volume segmentation,

we calculate and harmonise radiomics features and we propose the combination of tumour and lung lobes

radiomics features towards improving the classification performance. Having a rich feature set as a basis,

several combinations of feature selection and classification methods are tested and compared. Multiple

repetitions of cross-validation and external testing splits are used to reach robust manner. The proposed

method is trained and tested on an integrated dataset comprised of two open datasets (the NSCLC-Radiomics

and the NSCLC-Radiogenomics dataset from the Cancer Imaging Archive). It achieves average Precision and

Recall of 77.5% and 78.7% respectively, which could be further improved with a more extensive training set.

Therefore, this can be the basis for a prioritisation tool regarding lung cancer cases and detailed

staging/treatment decisions.

1 INTRODUCTION

Lung cancer is the leading cause of cancer-related

mortality for both males and females with the daily

deaths to be more than 2.5 times more than colorectal

cancer, the second most common non-gender specific

cancer, or more than breast, prostate, and pancreas

cancer-related deaths together (Siegel et al, 2022).

Primary or second-hand smoking, COPD, family

history, or exposure to carcinogens, such as asbestos,

cadmium or diesel fumes, are some of the risk factors

(Thandra et al, 2021). Early diagnosis will have a

great impact on the management of lung cancer

https://orcid.org/0000-0001-8605-8593

https://orcid.org/0000-0001-5613-652X

https://orcid.org/0000-0003-1573-8123

https://orcid.org/0000-0001-8915-6658

patients since it is found that the five-year survival

rates reach the 57% when the cancer is diagnosed in

its early stages (Raz et al, 2007).

In INCISIVE project (https://incisive-project.eu/),

we aim to address some major challenges in lung

cancer diagnosis and treatment, using Artificial

Intelligence tools and big data. Supporting and

automating lung cancer staging has been recognized

as one of the important challenges, which can

facilitate accurate diagnosis and support treatment

planning.

Specifically, non-small cell lung cancer is one of

the two main categories of lung cancer. The disease

Fotopoulos, D., Filos, D., Xinou, E. and Chouvarda, I.

Towards Lung Cancer Staging via ultipositional Radiomics and Machine Learning.

DOI: 10.5220/0011781500003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 4: BIOSIGNALS, pages 317-324

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)

317

stage can reveal information regarding the size of the

tumour if it has spread in parts of the body and it is

important information when planning what kind of

treatment is required. Staging is performed at the

initial diagnosis of the patient and at a second time

after the beginning of treatment, using the Tumour-

Node-Metastasis system (TNM) (Rami-Porta et al,

2009).

Cancer imaging is mainly used for diagnosis,

evaluation and treatment planning. In Lung cancer,

CT screening is recommended for the detection of

lung cancer but also as a screening for high-risk

populations. Imaging data are used for the evaluation

of disease severity with the TNM classification

scheme proposed (Amin et al, 2017). In addition, the

National Comprehensive Cancer Network (NCCN –

https://www.nccn.org) has proposed guidelines for

the selection of the appropriate therapy based on the

TNM classification and staging of the patients, and

thus the accurate staging of the cancer remains a

major clinical challenge. The staging procedure is

currently performed by experts through inspection

and assessment of physical exams, biopsy results and

imaging tests, which involve health costs, time effort,

and invasive methods.

Introducing digital tools to facilitate this

procedure, in terms of speed, cost, or accuracy, would

be of great benefit. In this direction, radiomic analysis

aims to extract characteristics of specific structures

found in medical images, leading to the quantitative

analysis of images. Radiomics features have already

been combined with machine learning methods to

detect malignancy in lung cancer (Anagnostopoulos

et al, 2022), while additional clinical features, such as

histopathological analysis results, have been used to

improve the success rates of the above algorithms.

A number of previous research efforts have

proposed methods to identify the stage of the patient

non-invasively, using biomarkers that are extracted

from medical images. Yu et al. (Yu et al., 2019)

implemented a machine learning algorithm for

radiomics-based prediction of the pathological stage

of lung cancer. They reported that their results were

promising, being able to predict the tumour stages

with high accuracy, especially for lung

adenocarcinoma type of cancer. Another paper on this

topic, from (Kasinathan and Jayakumar, 2022),

presents a cloud-based system and one of its

components is a classifier for staging. They report a

97% accuracy of the model in this task for images

automaticallysegmented. In (Ubaldi et al., 2021),

authors,report a machine learning pipeline that

utilizes open data and radiomic feature extraction for

histological and overall stage classification. They

approached stage classification as a binary problem

between stages I – II and achieved the best results

with a Random Forest (AUC = 0.72 ± 0.04) and

Support Vector Machine (0.84 ± 0.03) classifier.

Interestingly, they also mention that while using 2

datasets, one for training and another for testing, they

obtained better results when they used the 1st dataset

for training and the 2nd for testing, than with the

opposite order. They attributed this to the

misrepresentation of the two classes (stages I-II) for

the 2nd dataset. Indeed, the accuracy of the reference

information and the harmonization requirements may

increase the complexity of the problem.

The goal of our study is to employ radiomic

features, extracted from both healthy and pathological

tissue to develop a machine learning model for the

accurate staging of the lung cancer case. We present

a binary classification scheme, which classifies stage

I and II vs stage III and IV using lung CT imaging

data. We propose the use of tumourcharacteristics

combined with those of both lung lobes for the

characterization of staging. Upon full automation,

this can be a valuable decision support tool for first-

line diagnosis.

Figure 1: Overview of the analysis steps.

preprocessing lung masks

discard invalid data

merge datasests

feature

extraction

radiomics features for tumour

radiomics features for the two

lobes-lung volumes

normalised tumour radiomics

features by lung volume features

relative difference features

etween the two lobes

reject near zero variance

feature

selection

iterations and

final

selection

repeat the following in 100 runs

make a train-test split

ment trainin

data

Kruskal-Wallis for statistical

nificance

reject correlate

ply feature selection (RFE,

Boruta, Scad+L2) per method,

find features consistently chosen

>50% runs

model

building and

model

comparision

100 runs splits

each run CV each of the feature

sets

train/tune models (svm linear,

RF,nnet, dnn, glmnet) and

majority/stacked ensembles

calculate average model

performance metrics for

com

arison

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

318

2 METHODS

Cancer staging is originally a multiclass problem. In

this work, we reduced it to a 2-class problem. Our

proposed solution makes use of two publicly

available datasets. Thus, the need for harmonization

and the need to synthetically balance and augment the

two classes were two crucial points. An overview of

the proposed approach is presented in Figure 1. To

increase statistical robustness, a repetitive procedure

was chosen for feature selection, and consensus

features were selected. Following model training and

testing were also repeated multiple times, with

different training/testing splits for cross-validation

and external testing, to produce more stable results.

2.1 Data Description

The unified dataset used for the development of the

model is comprised of two datasets available in the

TCIA archive:

 The Radiomics dataset (Aerts et al, 2014). It

contains 422 cases of non-small cell lung

cancer (NSCLC). For each case, pre-treatment

CT scans, segmentations of ROIs of the images

and clinical data are included. A manual

delineation by a radiation oncologist of the 3D

volume of the primary gross tumour volume

("GTV-1") and selected anatomical structures

(i.e., lung) are available. The clinical variables

available included age, TNM stages, Overall

stage (inferred from TNM), gender, survival

and other. The overall stage variable includes

data belonging to stages: I, II, III.

 The Radiogenomics database. It contains 211

cases of NSCLC (Bakr et al, 2018). It includes

data belonging to classes I to IV. For each case,

CT images and tumour segmentations are

available, together with biological and clinical

data, including among other survival, age,

gender. In this work, lung volume

segmentations were not available, and

therefore we applied the lungmask automated

segmentation pipeline, based on deep learning

(Hofmanninger et al, 2020).

In this work, the stages are grouped in two classes, C0

(I and II subtypes), and C1 (III, and IV subtypes). The

rationale behind this choice is twofold: a) this

distinction reflects the severity and need for different

treatment options, and b) the multiclass problem

would require a much higher number of samples per

class, therefore the simplification to a two-class

problem can lead to a more robust and useful

approach. After rejecting problematic and incomplete

samples, the final unified dataset includes 434

samples: 126 of which from the Radiogenomics

database, and the rest from the Radiomics database.

C0 has 198 samples and C1 has 236 samples, which

include annotations for the tumour volume, left and

right lung lobes, and needed clinical information. The

percentage of stages represented is as follows: Stage

I:147, Stage II:51, Stage III:232, and Stage IV:4.

2.2 Radiomics Features

2.2.1 Calculation of Radiomics Features

We employed radiomics features for the quantitative

description of medical images. The pyradiomics

pipeline (van Griethuysen et al, 2017) was employed

for the calculation of radiomics features from the 3D

volumes, resulting in 1218 features. These

corresponded to features from the original images, the

Laplacian filtered images, and the Wavelet images,

including First Order Statistics, Shape-based (3D and

2D) descriptors, Gray Level Co-occurrence Matrix

(GLCM), Gray Level Run Length Matrix (GLRLM),

Gray Level Size Zone (GLSZM), Neighbouring Gray

Tone Difference Matrix and Gray Level Dependence

Matrix (GLDM).

Using the above-mentioned established pipeline,

we calculated radiomics features for the following

volumes: tumour volume, left lobe volume, right lobe

volume. These volumes were already segmented,

either manually or automatically, as mentioned in

section 2.1. The calculation of features on the

different volumes resulted in the Tu, LVR, and RVR

radiomics vectors, respectively.

2.2.2 Multi-Source Harmonization of

Radiomics Features

Harmonization at image level or feature level is a

necessary step for multisite analysis (Mali et al,

2021), but also analysis of data produced by

modalities of different vendors, to remove unwanted

variation when combining data across scanners and

sites. In the dataset used in this work, two sites and

multiple vendors were identified. The data originating

from vendors with very small representation were

rejected, as harmonization of these data could be

problematic. The chosen approach included the

harmonization of radiomics features with Combat

(Orlhac et al., 2022) method. Specifically, the steps

followed were:

 Harmonisation of data from same vendor in the

two databases (batch per database), which

Towards Lung Cancer Staging via ultipositional Radiomics and Machine Learning

319

incorporated most samples coming from the

two databases.

 Harmonisation with the remaining data from

other vendors in the two databases (batch per

vendor)

In each step, the Combat pipeline was applied to

the feature set, with batches defined as above, and the

type of volume (e.g. tumour, left lobe, right lobe) as a

confounder. Any non-harmonized features that

presented statistically significant differences between

batches after harmonization were removed, to avoid

any bias related to batch effects.

2.2.3 Feature Extraction

Based on the radiomics of the tumour and those of the

two lung lobes, new features were extracted, to

express the tumour in contrast to background, and the

differences between the two lobes. More specifically,

the feature vectors Tu, LVR, RVR, defined in section

2.2.1, were employed to calculate the normalized

tumour radiomics TuNo, which expresses the tumour

radiomics features Tu divided by the average between

the left and right volume radiomics features (LVR,

RVR). This is expected to normalize the tumour

radiomics values (tumour values with respect to

background values), decrease the inter-subject

variability, and improve the harmonization effort

(Escudero Sanchez et al, 2021). TuNo features were

calculated as in Eq 1.

TuNo =2* Tu/(LVR+RVR) (1)

In addition, the inter-lobe relative difference VRD

was calculated between the radiomics features

vectors RVR and LVR as:

VRD=2*

 − 

/(LVR+RVR)

(2)

The VRD feature vector is expected to introduce

information about the environment around the

tumour. We chose to use the whole lobe volumes

instead of a region around the tumour border, to

increase simplicity and support automated pipelines,

rather than options that involve human annotation.

Eventually, the feature vector set available for

feature selection includes the Tu, TuNo and VRD

features, i.e. the tumour features, the normalized

tumour features and the inter-lobe relative

differences.

2.3 Cancer Staging Models

The feature selection and classification model

methodology are described below. One important

point introduced in this work is the need to address

the problem of availability of a large number of

features, also correlated, in a dataset with comparable

dimension. To improve robustness, the procedure is

repeated multiple times. In each time, a different

training and external testing dataset are split, and

average behaviour among repetitions is eventually

considered.

2.3.1 Feature Selection

The challenge in this feature selection was the high

number of features, which are to some extent

correlated.

We considered as pre-processing steps: a)

removing linearly correlated features, b) removing

non statistically significant features based on

kruskall-wallis test (KW) with threshold 0.05/N,

(N=number of features).

Following, for the selection of the most

informative features, we considered three methods,

namely Recursive Feature Elimination (RFE), Boruta

method and SCAD-L2 method (Zeng and Xie, 2014).

Using the above methods, we introduced an

iterative procedure (100 iterations), which included in

each step the following actions:

 Formation of a new Training/testing set split

(80%)

 In the training set, application of the pre-

processing step for the removal from the

feature list of statistically unsignificant and

correlated features.

 Data Augmentation via SMOTE (Chawla et al,

2002) in the training set to balance the classes

and increase the data size

 Feature selection with one of the above

methods (RFE, Boruta, SCAD+L2).

Based on the result of the repeated feature selection

procedure, we introduced a voting mechanism to

filter-in the features that were consistently selected in

at least 50% of the iterations. These constitute our

final feature list.

2.3.2 Training Ml Models

These final feature sets were used as inputs in model

training. The classification models employed in this

work were: a) SVM with linear kernel, b) Random

Forest, c) generalized linear model via penalized

maximum likelihood (R package ‘glmnet’)

(Friedman, 2010), d) Stacked Autoencoder Deep

Neural Network (R package ‘deepnet’), e) a majority

voting model, f) an ensemble model based on

generalized linear model (glm) of the above

pretrained models. The train/test split was again

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

320

repeated 100 times, and in each repetition, the

following steps took place:

 Train/test set split (80%),

 Training data Augmentation via SMOTE,

 Classification models with internal 5-fold

cross-validation and hyperparameter

optimization,

 Test performance metrics in each repetition.

The average test set performance metrics among

the 100 repetitions was calculated and used for further

model comparison.

Figure 2: For two TuNo features (log sigma 3.0mm glszm

Large Area Emphasis and wavelet LLL first order Total

Energy), the distribution of values in the two classification

groups, taking into account the whole dataset. "log sigma

3” refers to features calculated after Laplacian filtering with

sigma=3, wavelet LLL refers to low-pass filtering in all

directions.

3 RESULTS

3.1 Selected Features

The procedure started with a large number of features.

The harmonization procedure rejected not well

harmonized features, to avoid the introduction of

unwanted batch bias. Following, after the generation

of the TuNo and VRD features based on tumour and

lung volumes, the number of Tu, TuNo, VRD features

entering the feature selection pipeline was 1418.

Effort was paid to end-up with a smaller number of

important features for model training.

In each feature selection cycle, the pre-processing

step (cross-correlation and KW test) rejected several

features and resulted in a range of around 400

statistically significant features, which constituted the

pool of features for feature selection by RFE or

Boruta or Scad+V2.

Following, based on the intermediate feature

selection sets, i.e. the features selected by each of the

three mechanisms in each of the 100 training set

repetitions described above, the consensus features

for each feature selection method was produced,

including the tumour (Tu), tumour normalized

(TuNo) and relative volume differences (VRD) types

of features. In RFE, 15 features were selected, 9 of

which were TuNo features, and 6 VRD features. The

majority (9/15) was wavelet features, and the rest

were log (based on the Laplacian filtered image). In

Boruta, 368 features were selected, 117/ 177 /74 in

Tu, TuNo and VRD types, respectively. The features

originated from original, log filtered and wavelet-

based images. The SCAD+L2 method resulted in 187

features, with only a small number of features coming

from original images, and 32/117 /38 in Tu, TuNo and

VRD types. Figure 2 depicts the distribution of values

per class for two features.

Overall, these are texture features in their

majority. Most of them belong to the normalized

tumour feature type (TuNo), and some more in the

relative volume difference type (VRD), will only a

few selected features from the initial tumour

radiomics (Tu). This supports the choice for the

“meta-features” introduced in this work. All relative

difference texture features show higher relative

difference in wavelet HHH texture feature values in

C0 than C1, i.e. higher relative difference in the two

lobes in the less metastatic stages. Most tumour

normalized texture values are lower in C0, showing a

clear difference between the tumour and background

in the less severe stages.

3.2 Classification Results

Table I presents test set performance metrics, as

median and quartiles of the test set performance

metrics, repeated 100 times with different train/test

set split. The median precision ranges between 75-

80%, which suggests that when C1 (more severe

class) is predicted, it is in general true, and the False

Towards Lung Cancer Staging via ultipositional Radiomics and Machine Learning

321

Positive is small. The recall is slightly lower (71-

79%), which suggests that there are a few False

Negatives, i.e. some C1 that are not identified, an

issue that needs improvement. The best precision-

recall case is found in the

SCAD+L2-RD method, with

both having values above 77%. It is worth noting that,

as identified by (Webb et al, 1993) and (Wu et al,

2020), an interrater variability exists in the domain

and the clinical staging accuracy and concordance

with pathological values also can improve. The

average balanced accuracy, and its standard

deviations for all classification schemes are presented

in Table II. It can be seen that RF classifier overall

outperforms other schemes.

Table 1: Average Performance Metrics (Median and 1

-3

Quartile) in the Test Set. C1: Positive Class.

Sen=Sensitivity, Spec=Specificity, Prec=Precision,

Rec=Recall, BA = Balanced Accuracy. Ens=ensemble

classifier.

Boruta RFE SCAD+L2

Perf RF Ens RF Ens RF Ens

Sen

78.72

70.218

2.98

76.6

71.818

1.38

74.47

70.217

8.72

71.28

65.967

4.47

78.72

72.348

0.85

78.72

72.348

2.98

Spec

71.79

66.677

6.92

69.23

64.107

4.36

78.21

71.798

2.05

76.92

74.368

2.05

74.36

69.237

6.92

71.79

66.677

6.92

Prec

76.47

74.007

9.17

71.967

8.05

79.49

76.68

83.72

79.07

76.098

2.61

77.55

75.518

0.85

76.7

73.758

0.12

Rec

78.72

70.218

2.98

76.6

71.818

1.38

74.47

70.217

8.72

71.28

65.967

4.47

78.72

72.348

0.85

78.72

72.348

2.98

77.42

73.458

0.00

76.68

73.287

8.79

76.57

73.338

0.85

74.6

70.817

8.36

77.49

74.148

0.85

77.49

74.167

9.60

74.6

71.96

76.81

72.79

71.16

75.48

75.27

72.078

0.17

73.6

70.687

7.63

75.37

72.977

8.89

74.71

71.637

8.09

The most important feature for the random forest

classification was in the TuNo type, and belonged to

the log filtered image features, expressing texture as

‘glszm Large Area High Gray Level Emphasis’. This

is a measure of the distribution of large area size

zones, with a greater value indicative of more larger

size zones and more coarse textures. In the TuNo

normalised version, a lower value in C0 class would

mean (as depicted in Fig 1), potentially relating also

to the size of the tumour. The most important features

per feature selection method are listed in the

supplementary section.

4 CONCLUSIONS

In the current study, a data-driven approach is

presented towards the development of a classification

model for lung cancer staging.

Radiomic features, applied on CT images during

initial diagnosis, from the tumour volume and the

lung lobe volumes, were selected following three

feature selection methods. These were combined and

used as input in a machine learning model. Most

features selected from each of the three feature

selection methods (RFE, Boruta, SCAD+L2) belong

to the Tumour normalized (TuNo) and relative

volume differences (VRD) types of features, which

shows the virtue of this multipositional radiomics

approach. This can be related to the findings of

(Escudero Sanchez et al, 2021) with respect to

increased robustness of texture features after

normalisation with normal tissue, although in our

case the tissue comes from the tumour organ

environment and cannot be classified as perfectly

healthy. The less severe class (C0) shows higher

relative difference in wavelet texture values among

the two lobes, and lower normalised tumour textural

characteristics. However, as mentioned by

(Demircioğlu, 2022), one cannot conclude with a

minimal number of radiomics as digital biomarkers,

because “Feature relevance in radiomics strongly

depends on the model used” and “Considering

features selected by a single model is misleading”.

Therefore, a more comprehensive approach will be

employed to conclude with the most important

features from multiple models as candidate

biomarkers. With a balanced accuracy of 75 % and a

F1-score 77.5%, the results are quite promising,

although there is still room for improvement.

Although the sample size of the combined dataset

was larger than the ones analysed in similar studies,

we are positive that a larger sample would be

preferable. Thus, we aim to retrain the model with

data from other available open datasets, but also data

collected as part of the INCISIVE project. As a result,

also given the expected heterogeneity of the data

collected from different clinical sites, special

attention will be paid to the improvement of

harmonization techniques, both regarding the raw

imaging data but also the harmonization of radiomic

features.

The novelty of this work compared to other

efforts, lies in the combined use of a unified dataset

from two sources, a set of enhanced features based on

the relative differences of the lungs' and tumour's

radiomic features and a repetitive data split/testing to

eliminate possible variation in the predictive

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

322

performance of the model. We strongly believe that

there is room for improvement, therefore we plan to

enrich the dataset by including clinical data, and

features that relate to TNM logic as well as combine

clinical and pathological staging features. Upon

availability of a larger dataset, additional

classification algorithms will be investigated as to

whether they improve the classification results,

before moving to a finer multiclass classification

scheme. Finally, the incorporation of a fairness and

an explainability component is among the necessary

future steps, to ensure better credibility of the

proposed system, and facilitate its validation from a

clinical perspective (or health expert’s) as well as its

deployment in a clinical environment.

Table 2: Average Performance Metrics for all

model combinations. In bold the best performances.

BA = Balanced Accuracy.

Feature set Model Mean

Std BA

RFE Linear SVM 73.78 3.81

RF 76.03 5.11

Dnn 62.22 5.98

Glmnet 73.82 4.08

Majority

ensemble 73.62 4.28

glm ensemble 74.43 5.13

Boruta Linear SVM 66.03 5.53

RF 74.26 5.03

Dnn 68.16 6.14

Glmnet 67.77 5.73

Majority

ensemble 70.80 4.68

glm ensemble 72.61 4.64

SCAD+L2 Linear SVM 68.84 4.60

RF 75.34 4.70

Dnn 69.39 6.00

Glmnet 70.64 4.81

Majority

ensemble 73.18 4.86

glm ensemble 74.78 4.41

ACKNOWLEDGEMENTS

This work was partly funded by EU H2020 project

INCISIVE under grant agreement No 952179.

Thanks to the INCISIVE consortium and especially

the clinical experts for highlighting important lung

cancer clinical challenges as targets for AI research.

REFERENCES

Aerts, H. J. W. L., Velazquez, E. R., Leijenaar, R. T. H.,

Parmar, C., Grossmann, P., Carvalho, S., Bussink, J.,

Monshouwer, R., Haibe-Kains, B., Rietveld, D.,

Hoebers, F., Rietbergen, M. M., Leemans, C. R.,

Dekker, A., Quackenbush, J., Gillies, R. J., & Lambin,

P. (2014). Decoding tumour phenotype by noninvasive

imaging using a quantitative radiomics approach.

Nature Communications, 5(1), 4006.

https://doi.org/10.1038/ncomms5006

Amin, M. B., Greene, F. L., Edge, S. B., Compton, C. C.,

Gershenwald, J. E., Brookland, R. K., Meyer, L., Gress,

D. M., Byrd, D. R., & Winchester, D. P. (2017). The

Eighth Edition AJCC Cancer Staging Manual:

Continuing to build a bridge from a population-based to

a more “personalized” approach to cancer staging: The

Eighth Edition AJCC Cancer Staging Manual. CA: A

Cancer Journal for Clinicians, 67(2), 93–99.

https://doi.org/10.3322/caac.21388

Anagnostopoulos, A. K., Gaitanis, A., Gkiozos, I.,

Athanasiadis, E. I., Chatziioannou, S. N., Syrigos, K.

N., Thanos, D., Chatziioannou, A. N., & Papanikolaou,

N. (2022). Radiomics/radiogenomics in lung cancer:

Basic principles and initial clinical results. Cancers,

14(7). https://doi.org/10.3390/cancers14071657

Bakr, S., Gevaert, O., Echegaray, S., Ayers, K., Zhou, M.,

Shafiq, M., Zheng, H., Benson, J. A., Zhang, W.,

Leung, A. N. C., Kadoch, M., Hoang, C. D., Shrager,

J., Quon, A., Rubin, D. L., Plevritis, S. K., & Napel, S.

(2018). A radiogenomic dataset of non-small cell lung

cancer. Scientific Data, 5(1), 180202.

https://doi.org/10.1038/sdata.2018.202

Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer,

W. P. (2002). SMOTE: Synthetic minority over-

sampling technique. The Journal of Artificial

Intelligence Research, 16, 321–357.

https://doi.org/10.1613/jair.953

Demircioğlu, A. (2022). Evaluation of the dependence of

radiomic features on the machine learning model.

Insights into Imaging, 13(1), 28. https://doi.org/

10.1186/s13244-022-01170-2

Escudero Sanchez, L., Rundo, L., Gill, A. B., Hoare, M.,

Mendes Serrao, E., & Sala, E. (2021). Robustness of

radiomic features in CT images with different slice

thickness, comparing liver tumour and muscle.

Scientific Reports, 11(1), 8262. https://doi.org/

10.1038/s41598-021-87598-w

Towards Lung Cancer Staging via ultipositional Radiomics and Machine Learning

323

Friedman J, Hastie T, Tibshirani R (2010). Regularization

Paths for Generalized Linear Models via Coordinate

Descent. Journal of Statistical Software, 33(1), 1-22.

URL https://www.jstatsoft.org/v33/i01/

Hofmanninger, J., Prayer, F., Pan, J., Röhrich, S., Prosch,

H., & Langs, G. (2020). Automatic lung segmentation

in routine imaging is primarily a data diversity problem,

not a methodology problem. European Radiology

Experimental, 4(1), 50. https://doi.org/10.1186/s41747-

020-00173-2

Kasinathan, G., & Jayakumar, S. (2022). Cloud-based lung

tumor detection and stage classification using deep

learning techniques. BioMed Research International,

2022, 4185835. https://doi.org/10.1155/2022/4185835

Mali SA, Ibrahim A, Woodruff HC, Andrearczyk V, Müller

H, Primakov S, Salahuddin Z, Chatterjee A, Lambin P.

(2021) Making Radiomics More Reproducible across

Scanner and Imaging Protocol Variations: A Review of

Harmonization Methods. Journal of Personalized

Medicine; 11(9):842.

https://doi.org/10.3390/jpm11090842

Orlhac, F., Eertink, J. J., Cottereau, A.-S., Zijlstra, J. M.,

Thieblemont, C., Meignan, M., Boellaard, R., & Buvat,

I. (2022). A guide to ComBat harmonization of imaging

biomarkers in multicenter studies. Journal of Nuclear

Medicine: Official Publication, Society of Nuclear

Medicine, 63(2), 172–179. https://doi.org/10.2967/

jnumed.121.262464

Rami-Porta, R., Crowley, J. J., & Goldstraw, P. (2009). The

revised TNM staging system for lung cancer. Annals of

Thoracic and Cardiovascular Surgery: Official Journal

of the Association of Thoracic and Cardiovascular

Surgeons of Asia, 15(1), 4–9.

Raz, D. J., Zell, J. A., Ou, S.-H. I., Gandara, D. R., Anton-

Culver, H., & Jablons, D. M. (2007). Natural history of

stage I non-small cell lung cancer: implications for

early detection. Chest, 132(1), 193–199.

https://doi.org/10.1378/chest.06-3096

Siegel, R. L., Miller, K. D., Fuchs, H. E., & Jemal, A.

(2022). Cancer statistics, 2022. CA: A Cancer Journal

for Clinicians, 72(1), 7–33. https://doi.org/10.3322/

caac.21708

Thandra, K. C., Barsouk, A., Saginala, K., Aluru, J. S., &

Barsouk, A. (2021). Epidemiology of lung cancer.

Contemporary Oncology (Poznan, Poland), 25(1), 45–

52. https://doi.org/10.5114/wo.2021.103829

Ubaldi, L., Valenti, V., Borgese, R. F., Collura, G.,

Fantacci, M. E., Ferrera, G., Iacoviello, G., Abbate, B.

F., Laruina, F., Tripoli, A., Retico, A., & Marrale, M.

(2021). Strategies to develop radiomics and machine

learning models for lung cancer stage and histology

prediction using small data samples. Physica Medica:

PM: An International Journal Devoted to the

Applications of Physics to Medicine and Biology:

Official Journal of the Italian Association of

Biomedical Physics (AIFB)

, 90, 13–22.

https://doi.org/10.1016/j.ejmp.2021.08.015

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

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

Fillion-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. https://doi.org/10.1158/0008-5472.can-17-

0339

Webb, W. R., Sarin, M., Zerhouni, E. A., Heelan, R. T.,

Glazer, G. M., & Gatsonis, C. (1993). Interobserver

variability in CT and MR staging of lung cancer.

Journal of Computer Assisted Tomography, 17(6),

841–846. https://doi.org/10.1097/00004728-19931100

0-00001

Wu, D. Y., Spangler, A. E., Vo, D. T., de Hoyos, A., &

Seiler, S. J. (2020). Simplified, standardized methods to

assess the accuracy of clinical cancer staging. Cancer

Treatment and Research Communications, 25(100253),

100253. https://doi.org/10.1016/j.ctarc.2020.100253

Yu, L., Tao, G., Zhu, L., Wang, G., Li, Z., Ye, J., & Chen,

Q. (2019). Prediction of pathologic stage in non-small

cell lung cancer using machine learning algorithm

based on CT image feature analysis. BMC Cancer,

19(1), 464. https://doi.org/10.1186/s12885-019-5646-9

Zeng, L., & Xie, J. (2014). Group variable selection via

SCAD-L2. Statistics, 48(1), 49–66. https://doi.org/

10.1080/02331888.2012.719513

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

324