Use of Radiomics in Low Dose Chest CT: A Proposal for a Phantom

Multi-Centric Study

Maria Irene Tenerani

1,2 a

, Silvia Arezzini

, Antonino Formuso

, Francesca Lizzi

2 b

Enrico Mazzoni

, Stefania Pallotta

3,4,5

, Alessandra Retico

2 c

, Camilla Scapicchio

2 d

Cinzia Talamonti

3,4,5

and Maria Evelina Fantacci

1,2 e

Department of Physics, University of Pisa, Pisa, Italy

National Institute for Nuclear Physics, Pisa, Italy

University of Florence, Department of Experimental and Clinical Biomedical Sciences ”Mario Serio”, Florence, Italy

Medical Physics Unit, AOU Careggi, Florence, Italy

National Institute for Nuclear Physics, Florence, Italy

Keywords:

Lung Cancer Screening, Low-Dose Computed Tomography, Computed Tomography Acquisition Protocol,

Phantom, Multi-Centric Study, Radiomics, Medical Data Sharing Platform.

Abstract:

Radiomics is a quantitative biomedical image analysis tool involving the mathematical extraction of image

features that can be used, particularly in oncology, to build predictive models based on artiﬁcial intelligence for

diagnosis and treatment outcome prediction. In Lung cancer screening via Low-Dose Computed Tomography

(LDCT), radiomics-based models could increase lung nodules detectability simplifying the implementation

of large-scale screening. However, their transposition into clinical practice is slowed by the instability that

radiomic feature values show in changes in CT image acquisition and reconstruction parameters. To build

more robust models, it is essential to conduct multi-centric radiomic studies leveraging the use of various types

of phantoms to overcome the challenges associated with patient data complexity. However, many difﬁculties

may arise related to both the image acquisition and reconstruction process and the extraction and analysis of

radiomic features. In this paper, from the results of a pilot study conducted with two phantoms, guidelines for

a multi-centric radiomic study on phantoms LDCTs are proposed, focusing on crucial aspects such as phantom

positioning, image acquisition and reconstruction protocol, and radiomic feature extraction pipeline. Finally, a

XNAT-based platform for data sharing and management, image quality control implementation and radiomic

feature extraction automation is proposed.

1 INTRODUCTION

Radiomics is a quantitative analysis tool based on

the assumption that biomedical images contain more

information than can be directly perceived by hu-

man vision. These additional data are obtained

through mathematical extraction of high-dimensional

radiomic features by considering images voxel by

voxel. Radiomics has enormous potential in devel-

oping precision medicine, particularly in cancer de-

tection, diagnosis, prognosis, and treatment evalua-

https://orcid.org/0009-0000-6230-7858

https://orcid.org/0000-0003-0900-0421

https://orcid.org/0000-0001-5135-4472

https://orcid.org/0000-0001-5984-0408

https://orcid.org/0000-0003-2130-4372

tion. The development of radiomics-based models,

for example in lung disease, could lead to improve-

ments in clinical workﬂow in diagnosis, prognosis,

management, follow-up, and monitoring of treatment

response. Indeed, numerous radiomics-based mod-

els and combined radiomics and Deep Learning (DL)

models have been developed for the detection and

classiﬁcation of pulmonary nodules and for predict-

ing or monitoring treatment response (Louis et al.,

2024; Frix et al., 2021). Their use in clinical decision

support systems could simplify the identiﬁcation of

nodules, mitigate the problems associated with small

lesions, ease the work of radiologists resulting in im-

proved accuracy of diagnosis, and thus facilitate the

implementation of large-scale lung cancer screening

programs on the at-risk population while simultane-

Tenerani, M. I., Arezzini, S., Formuso, A., Lizzi, F., Mazzoni, E., Pallotta, S., Retico, A., Scapicchio, C., Talamonti, C. and Fantacci, M. E.

Use of Radiomics in Low Dose Chest CT: A Proposal for a Phantom Multi-Centric Study.

DOI: 10.5220/0013377300003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 403-412

ISBN: 978-989-758-731-3; ISSN: 2184-4305

403

ously decreasing their costs (Saied et al., 2023). How-

ever, the translation of radiomics-based models to

clinical practice is complicated by the lack of repeata-

bility and robustness of Computed Tomography (CT)

derived radiomic features due to their dependence

on image acquisition and reconstruction parameters,

such as tube current [mA], tube voltage [kVp], expo-

sure [mAs], slice thickness, and voxel size (Shaﬁq-ul

Hassan et al., 2017; Traverso et al., 2018). Moreover,

the data analysis is strongly affected by the variability

in scanner models, speciﬁc clinical acquisition proto-

cols with different acquisition parameters, and recon-

struction settings that are often unavoidable in current

clinical practice. Large multi-centric studies are es-

sential to build more robust models. However, this

type of study is challenging as the collection of patient

data from different centers for centralized analysis is

complex, and sometimes impossible, for dosimetric,

legal, ethical, administrative, and technical reasons.

One way to overcome these difﬁculties is to employ

CT phantoms which, although not capable of fully

representing the extreme complexity of the human

anatomy, allow the extraction of plausible CT-derived

radiomic features and conduct repeatability studies

without having to consider the radiation dose deliv-

ered (Mackin et al., 2015). A preliminary study al-

ready conducted by acquiring CT images of two phan-

toms with two different scanners, different dose val-

ues, and numerous iterative reconstruction blending

levels, to study the repeatability and robustness of CT

radiomic features, revealed some acquisition-related

difﬁculties due to the variability in scanners, acqui-

sition protocols, and reconstruction settings (Scapic-

chio et al., 2024c; Scapicchio et al., 2024b; Tenerani

et al., 2024). Some of these difﬁculties were solved by

repeating the acquisitions, which, however, resulted

in the need for additional machine time, which was

already limited for CT scanners used daily in clinical

practice. Other difﬁculties, however, were not fully

resolved, leading to the exclusion of some data from

the study or to the need to resort to post-processing

techniques, which were also time-consuming. To im-

prove the efﬁciency of data collection and the qual-

ity of the dataset, it is important to deﬁne a standard

CT image acquisition and reconstruction protocol and

a well-deﬁned radiomic feature extraction pipeline.

Therefore, the aim of this paper is to outline the steps

of acquisition and reconstruction of phantom CT im-

ages, starting from the experience gained at the San

Luca hospital in Lucca, Italy, during the preliminary

study, and to deﬁne a standard procedure for the ex-

traction and analysis of radiomic features.

2 PHANTOMS

The phantoms considered for the multi-centric study

are a commercial phantom and a custom phantom, de-

veloped speciﬁcally to conduct radiomic studies, and

already employed in the preliminary study.

The Catphan-500® (The Phantom Laboratory,

NY, USA) (Mail, 2013) is a commercially available

phantom, commonly employed in clinical procedures

for quality control. It has a cylindrical shape with a di-

ameter of 20 cm and consists of four modules to study

several image properties at different contrast levels,

as can be seen in Fig. 1. Speciﬁcally, the CTP404

module, used in the preliminary radiomic study, in-

cludes seven cylindrical inserts of 15-mm diameter

and 25-mm thickness, made of different materials, i.e.

Acrylic, Polystyrene, LDPE, PMP, Air, Teﬂon and

Delrin and a vial of the same dimension which can

be ﬁlled with water, all embedded in a uniform back-

ground.

Figure 1: Illustration of the Catphan-500® phantom

model (Mail, 2013).

The custom phantom has an irregular elliptical

shape with axes measuring approximately 29 cm and

19 cm and the lower edge is cut to allow correct po-

sitioning within the CT system, also aided by the two

black markings on the phantom. Inside the phantom,

twenty ﬁve inserts are present embedded in a layer

of epoxy resin and made of different materials, tex-

tures, shapes and sizes in order to produce a wide

range of radiomic features values capable of mimick-

ing those of clinical CT images (Pallotta et al., 2020).

Six inserts are made of homogeneous materials such

as lung tissue, bone tissue and water and have a cylin-

drical or cubic shape. The three cubic inserts have

a side of approximately 9 mm, one cylinder has a

diameter of approximately 15 mm and height of 10

mm while the other two have a diameter and height

of approximately 9 mm. The other nineteen inserts,

also with cylindrical or cubic shape and with differ-

ent ﬁlling percentages, were fabricated employing a

3D printer using PLA, FLEX, and PETG with paral-

lel, triesacube, triangle, 1/4 cube and gyroid patterns.

BIOIMAGING 2025 - 12th International Conference on Bioimaging

404

The cubic inserts have a side of approximately 17 mm

while the cylindrical ones have a diameter and height

of approximately 17 mm. The exact structure of the

phantom is shown in Figure 2 and 3.

(a) (b)

Figure 2: Pictures of the custom phantom in the frontal (a)

and transverse (b) planes.

Figure 3: Custom phantom structure with the different in-

serts divided by material, ﬁlling percentages and texture

types.

3 PRELIMINARY RADIOMIC

STUDY

3.1 Acquisition Procedure

In the preliminary baseline study, two different com-

mercial CT scanners, available at the San Luca hos-

pital, Lucca, Italy, were employed to acquire the CT

images of the two phantoms: the Revolution Evo 64

Slice scanner (GE Healthcare) and the Aquilon CX

128 Slice CT scanner (Toshiba). An initial dataset

of the Catphan® phantom was acquired by starting

with the institutional clinical protocol for diagnostic

tasks in chest imaging and exploring four Computed

Tomography Dose Index (CDTI

vol

) values (IAEA,

2012), ranging from ultra-low to high dose, and four

iterative reconstruction blending levels, for a total

of thirty-two different protocols, each repeated three

times, and ninety-six CT scans. The second custom

phantom dataset consisted of thirty CT images ac-

quired with ten protocols considering two CT scan-

ners, three CDTI

vol

values (from ultra-low-dose to

standard dose), four iterative blending levels, and re-

peating the acquisition three times for each set of pa-

rameters. The acquisition parameters used to produce

the two datasets are described in the Table 1 and Ta-

ble 2.

3.2 Features Repeatability and

Robustness

Eighty-six ﬁrst- and second-order features were ex-

tracted from each of the homogeneous inserts present

in the CTP404 Catphan® phantom module and from

each of the nineteen textured inserts present in the

custom phantom directly from the original images,

that is without applying any ﬁlter. In particular,

radiomic features belonging to ﬁrst-order statistics,

Gray Level Co-occurrence Matrix (GLCM), Gray

Level Size Zone Matrix (GLSZM), Gray Level Run

Length Matrix (GLRLM) and Gray Level Depen-

dence Matrix (GLDM) classes were calculated using

Pyradiomics. PyRadiomics is a ﬂexible open-source

Python package that enables the processing and the

extraction of a large number of radiomic features

from both 2D and 3D medical images in compliance

with feature deﬁnitions as described by the Imaging

Biomarker Standardization Initiative (IBSI) (Van Gri-

ethuysen et al., 2017; Zwanenburg et al., ).

The statistical analysis of the radiomic features

extracted from the Catphan® dataset showed that im-

age quality, assessed by calculating the Detectability

Index (Samei et al., 2019) on the polystyrene insert,

inﬂuences the robustness of the radiomic features,

quantiﬁed using the two-way mixed effect model with

average raters type and absolute agreement Intraclass

Correlation Coefﬁcient (ICC) (McGraw and Wong,

1996; Koo and Li, 2016). Speciﬁcally, a greater per-

centage of radiomic features extracted from the var-

ious inserts are found to be robust when considering

images, and thus protocols, that have similar image

quality than when considering images with widely

varying image quality; In fact, it was already ver-

iﬁed that, when considering protocols with similar

image quality, about 80% of the features extracted

from the polystyrene insert were found to be ro-

bust, where robust features are those that have an

ICC≥75% (Scapicchio et al., 2024c).

This behavior was also conﬁrmed for the textured

inserts present within the custom phantom, where

about 80% robust radiomic features were obtained

for the more homogeneous inserts, while, for the

more deﬁned textured inserts, the percentage of ra-

diomic features considered robust dropped signiﬁ-

cantly, down to approximately 20% (Tenerani et al.,

2024). The percentages of robust features for the

Use of Radiomics in Low Dose Chest CT: A Proposal for a Phantom Multi-Centric Study

405

Table 1: Acquisition and reconstruction parameters of the thirty-two protocols used to acquire the Catphan-500® CT images

with the two CT scanners.

Revolution GE Aquilon Toshiba

CTDI

vol

[mGy] (Tube current [mA])

High 13.52 (160) 16.50 (300)

Standard 6.76 (80) 8.30 (150)

Reduced 4.06 (50) 5.00 (90)

Low 2.03 (25) 2.49 (45)

Data acquisition

Tube potential (kVp) 120 120

Pitch 0.984 0.938

Image Reconstruction

Pixel Spacing (mm) 0.406 0.427

Slice thickness (mm) 1.25 1.00

Kernel LUNG FC56

Iterative level 0%, 10%, 40%, 70% 0%, mild, standard, strong

Table 2: Acquisition and reconstruction parameters of the ten protocols used to acquire the custom phantom CT images with

the two CT scanners. The pure Filtered-Back-Projection reconstruction was not applied to the Toshiba low dose acquisitions.

Revolution GE Aquilon Toshiba

CTDI

vol

[mGy] (Tube current [mA])

Standard 7.1 (80) /

Reduced / 4.4 (40)

Low / 2.2 (20)

Data acquisition

Tube potential (kVp) 120 120

Pitch 0.984 0.938

Image Reconstruction

Pixel Spacing (mm) 0.703 0.781

Slice thickness (mm) 1.25 1.00

Kernel LUNG FC56

Iterative level 0%, 10%, 70% 0%*, mild, standard, strong

Catphan® phantom polystyrene insert and the custom

phantom inserts are shown in Table 3. Repeatability

tests conducted with the custom phantom also showed

how the positioning of the phantom and the radiomic

features extraction pipeline, particularly the deﬁnition

of the Regions Of Interest (ROIs) from which fea-

tures are extracted, affects their value and thus their

repeatability. Indeed, shifts of a few voxels in the def-

inition of the ROIs (about two in each direction) led

to a percentage of non-repeatable features up to about

25% for some inserts while changes in the volume of

the ROIs caused a percentage of non-repeatable fea-

tures ranging from about 30% to 50%, depending on

the insert considered.

4 DISCUSSION OF

PRELIMINARY WORK AND

COMPARISON WITH

LITERATURE

Although there are some limitations in the prelimi-

nary study, such as the employment of phantoms only,

the limited number of protocols used for the cus-

tomized phantom acquisitions, and the difﬁculties re-

lated to positioning the phantoms within the gantry,

the results described highlight the need to establish

a standard protocol for image acquisition and recon-

struction to collect data from the various clinical cen-

ters in the context of a multi-centric study. It is there-

fore critical to identify the physical acquisition pa-

rameters such as tube current, tube potential, pitch,

slice thickness and pixel spacing and the ideal itera-

tive reconstruction blending level to reduce the radi-

BIOIMAGING 2025 - 12th International Conference on Bioimaging

406

Table 3: Percentage of robust features extracted from the polystyrene insert of the Catphan® phantom (Scapicchio et al.,

2024c) and on a more homogeneous insert and a more deﬁned textured insert of the custom phantom (Tenerani et al., 2024).

Insert % of robust features

Polystyrene (Catphan®) ∼ 80%

More homogeneous insert (custom phantom) ∼ 80%

More deﬁned textured insert (custom phantom) ∼ 20%

ation dose delivered as much as possible while con-

sidering the different behavior of various commercial

iterative algorithms in suppressing noise and dealing

with interfaces (Samei et al., 2019).

In order to develop and validate a radiomic model

that can be implemented within lung cancer screening

programs, it is essential to use CT acquisition proto-

cols that are as consistent as possible with those com-

monly used in clinical practice. Although several CT

scanners and low-dose CT acquisition protocols have

been used in large lung cancer screening studies, such

as NLST, NELSON, UKLS, and ITALUNG (Team,

2011; Zhao et al., 2011; Baldwin et al., 2011; Pegna

et al., 2009), in which speciﬁc reconstruction values

are often not clearly reported (Vonder et al., 2021),

currently more attention is being paid to the choice

of protocols. Indeed, the American College of Ra-

diology (ACR), the Society of Thoracic Radiology

(STR) and the European Society of Thoracic Imag-

ing (ESTI) have provided several guidelines for the

choice of CT acquisition and reconstruction protocols

in lung cancer screening (American College of Ra-

diology (ACR), 2023; European Society of Thoracic

Imaging (ESTI), 2019). In addition, the American

Association of Physicists in Medicine (AAPM) de-

veloped a set of detailed acquisition and reconstruc-

tion protocols of over 30 CT systems of six major

vendors for lung cancer screening purposes (Ameri-

can Association of Physicists in Medicine (AAPM),

2023). Another aspect to consider when planning CT

acquisitions for lung cancer screening is image qual-

ity, which directly affects the detectability of lung

nodules that can be very small, on the order of a few

millimeters, and difﬁcult to detect within the lung tis-

sue (Thakur et al., 2020). The Quantitative Imag-

ing Biomarker Alliance (QIBA) provides six stan-

dard markers for image quality assurance, i.e., mini-

mum requirements for image quality deﬁned by reso-

lution, edge enhancement, HU deviation, voxel noise,

and spatial image distortion (Quantitative Imaging

Biomarkers Alliance (QIBA), 2018). To investigate

the adherence of the protocols suggested by AAPM

to the quality standards proposed by QIBA, in the

study by Iball et al. the use of two different recon-

struction kernels, a sharper kernel as suggested by the

speciﬁc AAPM protocol, and a smoother kernel, is

investigated. The authors evaluated the six metrics

suggested by the QIBA guidelines for image qual-

ity using the CTLX1 phantom and found that when

imaging the phantom using the AAPM scan proto-

col with the suggested sharp kernel, the image qual-

ity failed the QIBA speciﬁcation for two of the six

metrics while, using a smoother kernel, all six image

quality speciﬁcations were met (Iball et al., 2021).

These results highlight the importance of the choice

of the reconstruction kernel in deﬁning the acquisi-

tion protocol with regard to both the detectability of

lung nodules and the robustness of radiomic features.

To select a deﬁned CT protocol, it is possible

to start from the guidelines, particularly the individ-

ual CT scanner-speciﬁc parameters proposed in the

AAPM guide, then explore also a reduced dose value

and reconstruct the image with different kernels (soft

and sharp) and at least two different iterative recon-

struction blending levels, similar to what was done

for the Reduced and Low Dose protocols in the pre-

liminary study phantom datasets.

Given the limited machine time usually available

to perform these acquisitions when using CT scanners

employed in daily clinical practice, it is of paramount

importance to establish guidelines on the placement

of the phantom within the scanner to make the image

acquisition process repeatable and efﬁcient. The im-

portance of the phantom positioning step inside the

gantry is accentuated by the different shape of the

patient couch proper to the CT scanners; in fact, in

some scanners, typically for diagnostic use, the couch

is concave while for others, typically used for center-

ing in radiotherapy, the bed is ﬂat. Once the scanning

is done, the quality of image acquisition and recon-

struction should be checked quickly so that it would

be possible, in case of discrepancies, to reconstruct or

reacquire the CT scans immediately.

Another essential aspect to consider, in view of

the expansion of this study to more centers, concerns

the identiﬁcation of a deﬁned pipeline for feature ex-

traction that minimizes ROIs mismatch problems; in

fact, a further weakness of the preliminary study was

precisely related to the difﬁculties of positioning the

extraction ROIs within the inserts due to small and

sometimes difﬁcult-to-avoid differences in the posi-

tioning of the phantoms in different scanners. This

could be accomplished by exploring different feature

extraction parameters, such as bin width (Larue et al.,

Use of Radiomics in Low Dose Chest CT: A Proposal for a Phantom Multi-Centric Study

407

2017), and by developing an automatic image coregis-

tration and ROIs placement system that can be easily

replicated on images from different sites. This is es-

pecially important considering that images acquired

on different scanners will have different pixel spac-

ing and slice thickness values, and it may therefore be

necessary to resample the images to isotropic voxels

using interpolations. This aspect is even more pro-

nounced when considering scans of patients where

segmentation of, for example, lung nodules is often

critical. Evaluation of the repeatability of radiomic

features with respect to changes in extraction ROIs,

even if in simpliﬁed structures such as those found

within phantoms, could provide insights regarding

how to deal with different segmentations in real pa-

tient scans.

On the basis of the elaborated considerations, we

propose as future work a multi-centric phantom study

for which we suggest some guidelines regarding the

positioning of the phantoms within the CT scanner

gantry, the implementation of quality control checks

on the acquired images, the extraction and processing

pipeline of radiomic features, and we suggest the use

of a dedicated data storage and sharing XNAT-based

platform.

5 GUIDELINES FOR EFFICIENT

DATA COLLECTION

5.1 Phantom Positioning Step

The ﬁrst key step in conducting a multicenter study

is to ensure the proper positioning of the phantom

within the gantry of the CT scanner. For the com-

mercial Catphan-500® phantom, speciﬁc guidelines

for its positioning and alignment can be found within

the manual (Mail, 2013) and an example is shown

in Fig. 4. In contrast, for the custom phantom, no

guidelines are available yet. Here, a procedure is

proposed for the correct and reproducible placement

of the custom phantom that could be easily replicated

with the different CT scanners. In CT scanners

equipped with a ﬂat couch, the custom phantom

can simply be positioned with its base using a level,

while, for CT scanners equipped with a concave

couch, a wooden slab can be used and placed on

the couch and aligned with a level above which

the custom phantom is positioned. To ensure the

correct alignment of the custom phantom with

the center of the imaging system, one positioning

laser of the CT scanner must be aligned with the align-

ment marker marked in the direction of the height of

Figure 4: Catphan-500® phantom acquisition setting. The

phantom is placed on its case leveled and aligned with the

scanner alignment markers.

Figure 5: Custom phantom acquisition setting. The phan-

tom is placed with its base on the wooden slab leveled and

aligned with the scanner alignment markers.

the phantom, while the other positioning laser must

be aligned with the edge of the phantom face itself.

5.2 Quality Control Checks

Inconsistencies in data acquisition and reconstruction

parameters may compromise the stability of radiomic

features and thus deteriorate the generalizability of

radiomics-based models. Improperly acquired im-

ages often require the use of additional image post-

processing or to repeat the reconstruction process thus

requiring additional operator time. In cases where er-

rors in acquisition parameters have occurred, it is nec-

essary to either remove the speciﬁc acquisition from

the dataset or repeat it, with relative difﬁculty in al-

locating additional machine time. Therefore, it is es-

sential to monitor the acquisition and reconstruction

parameters as soon as possible to detect any discrep-

ancies from the chosen protocol and be able to either

reconstruct or repeat the image immediately. It is nec-

essary to check both that the acquisition parameters,

i.e., tube voltage, tube current, activation or deacti-

BIOIMAGING 2025 - 12th International Conference on Bioimaging

408

Figure 6: Feature extraction ROIs on the polystyrene insert

of the Catphan® phantom.

vation of Automatic Exposure Control (AEC), pitch

and voxel size, and reconstruction parameters, e.g.,

image reconstruction kernel and iterative reconstruc-

tion blending level, correspond to those deﬁned in the

chosen protocol for the speciﬁc CT scanner. In fact,

difﬁculties in reconstructing images with a speciﬁc it-

erative reconstruction blending level and ﬁxed kernel

may happen when the CT scanner is set to acquire

many subsequent acquisition, as noted in the prelimi-

nary study. It is also important to verify that the posi-

tioning and alignment of the phantom are as expected,

that no reconstruction artifacts are present in the im-

age and that the correct number of repetitions were

acquired.

6 FEATURE EXTRACTION

PIPELINE

Since the CT images will be acquired at different clin-

ical centers using many different CT scanners, even

from different vendors, in order to perform rigid reg-

istration of the images, it is important to resample the

images to isotropic 1 × 1 × 1 mm

voxel using cubic

spline interpolation and nearest neighbor interpola-

tion and to convert images intensities to Hounsﬁeld

units (HU) (Louis et al., 2024). Subsequently, the Re-

gions Of Interests (ROIs) from which radiomic fea-

tures will be extracted on the cylindrical inserts of the

Catphan® and those in the custom phantom should be

placed. For the Catphan® phantom, it is sufﬁcient to

choose the cylindrical ROIs centered in the center of

the insert with a distance of about 2 voxels in each

direction from the edge of the insert, as shown in Fig-

ure 6. For the custom phantom, on the other hand, it

is important to keep a greater number of voxels from

the edge of the insert for both the cylindrical and the

cubic ROIs, at least ﬁve, because of the presence of a

capsule inside which the insert is contained and resid-

ual glue with which the inserts were glued to a sheet

of plexiglass to anchor them during the construction

of the phantom itself, as shown in Figure 7.

Figure 7: Feature extraction ROIs positioned on the custom

phantom cylindrical and cubic inserts.

Once the ROIs have been deﬁned, it is possible to

proceed with the extraction of ﬁrst- and second-order

features, excluding shape features, using the Pyra-

diomics open-source python package. PyRadiomics

offers several settings to customize feature extrac-

tion, including the bin width for image gray level dis-

cretization. The default value of this parameter is

set to 25 HU, which can be used as a starting value

for feature extraction, after which two other values,

such as 10 HU and 50 HU, can be used to explore the

dependence of radiomic feature values on bin width.

Of the extracted radiomic features, repeatability and

robustness will then be evaluated through the assess-

ment of speciﬁc metrics such as the Intraclass Corre-

lation Coefﬁcient (ICC) and the Coefﬁcient of Vari-

ation (CV). The signiﬁcance of the robust features

will instead be evaluated through the development of

classiﬁers either based on Machine Learning and with

the implementation of hybrid models combining ra-

diomic features and Deep Learning. The developed

hybrid classiﬁer could then be translated to chest CT

images for lung nodule classiﬁcation with the aim of

lowering the amount of false positives, an essential

aspect for the feasibility of large-scale lung cancer

screening.

7 DATA SHARING PLATFORM

Phantom CT images collected by the various clinical

centers could be stored within a platform based on the

Extensible Neuroimaging Archive Toolkit (XNAT)

that is already under development (Scapicchio et al.,

2024a). XNAT is an open source software platform

developed to support FAIR principles, which is a set

of guidelines to ensure the Findability, Accessibil-

ity, Interoperability, and Reusability of data, and to

facilitate the management of medical data. It was

initially developed to store and share neuroimaging

data and was later extended to other areas of medi-

cal imaging (Marcus et al., 2007; Tim

on et al., 2017).

XNAT-based platforms can be conﬁgured in a vari-

ety of ways to optimize project data management.

Use of Radiomics in Low Dose Chest CT: A Proposal for a Phantom Multi-Centric Study

409

Figure 8: Display of a central slice of the custom phantom

in the OHIF-XNAT viewer.

Data can be uploaded directly in DICOM format and

the images could be visualized through the integrated

OHIF viewer, as shown in Figure 8, as well as orga-

nized, shared, searched and downloaded on the plat-

form while also implementing an access management

mechanism that would allow various clinical centers

to login to the platform. The ability to store and

share heterogeneous data makes the XNAT technol-

ogy effective in multi-centric data storage as it al-

lows interconnected data to be managed across dif-

ferent projects. An added value of using a platform

based on XNAT is the possibility to directly imple-

ment an image quality control pipeline, i.e., an auto-

mated quality control analysis tool on the CT acqui-

sition that would be executed within the platform as

soon as these are uploaded on the platform. Another

potential of this platform concerns the development

of innovative integrated plugins to perform external

analysis such as, for example, the possibility of eval-

uating standard image quality metrics, i.e., Contrast

to Noise Ratio (CNR), Resolution and Noise Power

Spectrum (NPS), but also more complex metrics such

as the Detectability Index. Automated calculation of

these metrics would provide an immediate indication

of the quality of the acquired images by highlighting

the presence of any acquisition, reconstruction or po-

sitioning issues.

The XNAT platform could also be used for the au-

tomation of the radiomic pipeline through the ability

to use OHIF viewer tools for the automatic contour-

ing of ROIs. This would speed up, simplify and auto-

mate the process of deﬁning ROIs within inserts and

extracting features, improving the repeatability of ra-

diomic analysis, a critical aspect of multi-centric stud-

ies. Therefore, the use of the XNAT platform would

not be limited to data storage, but could also become a

tool to optimize the steps of image acquisition and re-

construction parameter control, image quality assess-

ment, and radiomic feature extraction pipeline that

each clinical center could access.

8 CONCLUSIONS

Radiomic-based models and Radiomics and Deep

Learning combined models have huge potential with

regard to aiding diagnosis for lung nodule detection,

particularly in lung cancer screening programs where

a large number of chest LDCT must be analyzed by

detecting even very small lung nodules. However, the

translation to clinical practice of these models is lim-

ited by the poor reproducibility of radiomic features

as the image acquisition and reconstruction parame-

ters and the pipeline of radiomic feature extraction it-

self vary. Multi-centric phantom studies are essential

to enable harmonization of data acquired with differ-

ent CT scanners, deﬁne a procedure to identify a sub-

set of stable radiomic features using phantoms speciﬁ-

cally developed for radiomics study that could be gen-

eralized to more heterogeneous datasets, and evaluate

the reliability of the results obtained in clinical tri-

als. In this study, a set of guidelines are proposed

for conducting a multi-centric study with two phan-

toms, one commercial and one custom. Particular at-

tention should be paid to the choice of image acqui-

sition and reconstruction protocols, starting with the

standard for lung nodule detection, and to the posi-

tioning of the phantoms. Image quality checks must

be implemented to verify adherence between the ac-

quired images and the acquisition and reconstruction

parameters required by the protocol. Finally, a spe-

ciﬁc feature extraction pipeline must be followed to

allow for reproducibility. The data should be stored

within a platform that allows the sharing among the

various centers. The proposed XNAT-based platform

could facilitate data management by simplifying the

image quality control procedure and automating the

radiomic feature extraction pipeline.

ACKNOWLEDGMENTS

The research leading to these results has received

funding from:

The European Union - NextGenerationEU through

the Italian Ministry of University and Research

under: PNRR - M4C2-I1.3 Project PE 00000019

”HEAL ITALIA” to Maria Evelina Fantacci and

Maria Irene Tenerani – CUP I53C22001440006.

Piano Nazionale di Ripresa e Resilienza

(PNRR), Missione 4, Componente 2, Ecosis-

temi dell’Innovazione–Tuscany Health Ecosystem

(THE), Spoke 1 “Advanced Radiotherapies and

Diagnostics in Oncology”—CUP I53C22000780001.

PNRR - M4C2 - Investimento 1.3, Partenariato

Esteso PE00000013 - ”FAIR - Future Artiﬁcial

BIOIMAGING 2025 - 12th International Conference on Bioimaging

410

Intelligence Research” - Spoke 8 ”Pervasive AI”,

funded by the European Commission under the

NextGeneration EU programme.

PNRR - M4C2 - I1.4, CN00000013 - ”ICSC –

Centro Nazionale di Ricerca in High Performance

Computing, Big Data and Quantum Computing”

- Spoke 8 ”In Silico medicine and Omics Data”,

both funded by the European Commission under the

NextGeneration EU programme.

The National Institute for Nuclear Physics (INFN)

within the next AIM (Artiﬁcial Intelligence in

Medicine: next steps) research project (INFN-

CSN5), https://www.pi.infn.it/aim.

The views and opinions expressed are those of the

authors only and do not necessarily reﬂect those of

the European Union or the European Commission.

Neither the European Union nor the European Com-

mission can be held responsible for them.

REFERENCES

American Association of Physicists in Medicine (AAPM)

(2023). Lung Cancer Screening CT Protocols Version

6.0. AAPM. Accessed online.

American College of Radiology (ACR) (2023).

ACR–SABI–SPR–STR Practice Parameter for

the Performance of Thoracic Computed Tomography

(CT). American College of Radiology. Practice

parameters for thoracic CT imaging.

Baldwin, D., Duffy, S., Wald, N., Page, R., Hansell, D.,

and Field, J. (2011). Uk lung screen (ukls) nodule

management protocol: modelling of a single screen

randomised controlled trial of low-dose ct screening

for lung cancer. Thorax, 66(4):308–313.

European Society of Thoracic Imaging (ESTI) (2019).

Chest CT for Lung Cancer Screening: ESTI Techni-

cal Standards. ESTI. Technical standards for lung

cancer screening CT.

Frix, A.-N., Cousin, F., Refaee, T., Bottari, F.,

Vaidyanathan, A., Desir, C., Vos, W., Walsh, S., Oc-

chipinti, M., Lovinfosse, P., et al. (2021). Radiomics

in lung diseases imaging: state-of-the-art for clini-

cians. Journal of Personalized Medicine, 11(7):602.

IAEA (2012). Quality assurance programme for com-

puted tomography: Diagnostic and therapy applica-

tions. IAEA Human Health Series, 19.

Iball, G. R., Darby, M., Gabe, R., Crosbie, P. A., and Callis-

ter, M. E. (2021). Establishing scanning protocols for

a ct lung cancer screening trial in the uk. The British

journal of radiology, 94(1128):20201343.

Koo, T. K. and Li, M. Y. (2016). A guideline of selecting

and reporting intraclass correlation coefﬁcients for re-

liability research. Journal of chiropractic medicine,

15(2):155–163.

Larue, R. T., van Timmeren, J. E., de Jong, E. E., Feliciani,

G., Leijenaar, R. T., Schreurs, W. M., Sosef, M. N.,

Raat, F. H., van der Zande, F. H., Das, M., et al.

(2017). Inﬂuence of gray level discretization on ra-

diomic feature stability for different ct scanners, tube

currents and slice thicknesses: a comprehensive phan-

tom study. Acta oncologica, 56(11):1544–1553.

Louis, T., Lucia, F., Cousin, F., Mievis, C., Jansen, N.,

Duysinx, B., Le Pennec, R., Visvikis, D., Neb-

bache, M., Rehn, M., et al. (2024). Identiﬁca-

tion of ct radiomic features robust to acquisition and

segmentation variations for improved prediction of

radiotherapy-treated lung cancer patient recurrence.

Scientiﬁc Reports, 14(1):9028.

Mackin, D., Fave, X., Zhang, L., Fried, D., Yang, J., Tay-

lor, B., Rodriguez-Rivera, E., Dodge, C., Jones, A. K.,

et al. (2015). Measuring computed tomography scan-

ner variability of radiomics features. Investigative ra-

diology, 50(11):757–765.

Mail, T. B. (2013). Catphan® 500 and 600 manual. The

Phantom Laboratory.

Marcus, D. S., Olsen, T. R., Ramaratnam, M., and Buckner,

R. L. (2007). The extensible neuroimaging archive

toolkit: an informatics platform for managing, explor-

ing, and sharing neuroimaging data. Neuroinformat-

ics, 5:11–33.

McGraw, K. O. and Wong, S. P. (1996). Forming inferences

about some intraclass correlation coefﬁcients. Psycho-

logical methods, 1(1):30.

Pallotta, S., Cusumano, D., Taddeucci, A., Benelli, M.,

Sulejmeni, R., Lenkowicz, J., Calusi, S., Marrazzo,

L., Talamonti, C., Belli, G., et al. (2020). Po-1536:

Radiomik: a phantom to test repeatability and repro-

ducibility of ct-derived radiomic features. Radiother-

apy and Oncology, 152:S830–S831.

Pegna, A. L., Picozzi, G., Mascalchi, M., Carozzi, F. M.,

Carrozzi, L., Comin, C., Spinelli, C., Falaschi, F.,

Grazzini, M., Innocenti, F., et al. (2009). Design, re-

cruitment and baseline results of the italung trial for

lung cancer screening with low-dose ct. Lung cancer,

64(1):34–40.

Quantitative Imaging Biomarkers Alliance (QIBA) (2018).

QIBA Proﬁle: Small Lung Nodule Volume Assessment

and Monitoring in Low Dose CT Screening.

Saied, M., Raafat, M., Yehia, S., and Khalil, M. M. (2023).

Efﬁcient pulmonary nodules classiﬁcation using ra-

diomics and different artiﬁcial intelligence strategies.

Insights into Imaging, 14(1):91.

Samei, E., Bakalyar, D., Boedeker, K. L., Brady, S., Fan, J.,

Leng, S., Myers, K. J., Popescu, L. M., Ramirez Gi-

raldo, J. C., Ranallo, F., et al. (2019). Performance

evaluation of computed tomography systems: sum-

mary of aapm task group 233. Medical physics,

46(11):e735–e756.

Scapicchio, C., Arezzini, S., Fantacci, M., Formuso, A.,

Kraan, A., Mazzoni, E., Saponaro, S., Tenerani, M.,

and Retico, A. (2024a). Integration and optimiza-

tion of xnat-based platforms for the management of

heterogeneous and multicenter data in biomedical re-

search. In Proceedings of the 13th International Con-

ference on Data Science, Technology and Applications

- DATA, pages 551–558. INSTICC, SciTePress.

Use of Radiomics in Low Dose Chest CT: A Proposal for a Phantom Multi-Centric Study

411

Scapicchio, C., Imbriani, M., Lizzi, F., Quattrocchi, M.,

Retico, A., Saponaro, S., Tenerani, M. I., Tofani,

A., Zafaranchi, A., and Fantacci, M. E. (2024b).

Characterization and quantiﬁcation of image qual-

ity in ct imaging systems: A phantom study. In

Proceedings of the 17th International Joint Confer-

ence on Biomedical Engineering Systems and Tech-

nologies - BIOIMAGING, pages 289–296. INSTICC,

SciTePress.

Scapicchio, C., Imbriani, M., Lizzi, F., Quattrocchi, M.,

Retico, A., Saponaro, S., Tenerani, M. I., Tofani, A.,

Zafaranchi, A., and Fantacci, M. E. (2024c). Inves-

tigation of a potential upstream harmonization based

on image appearance matching to improve radiomics

features robustness: a phantom study. Biomedical

Physics & Engineering Express, 10(4):045006.

Shaﬁq-ul Hassan, M., Zhang, G. G., Latiﬁ, K., Ullah, G.,

Hunt, D. C., Balagurunathan, Y., Abdalah, M. A.,

Schabath, M. B., Goldgof, D. G., Mackin, D., et al.

(2017). Intrinsic dependencies of ct radiomic features

on voxel size and number of gray levels. Medical

physics, 44(3):1050–1062.

Team, N. L. S. T. R. (2011). The national lung screen-

ing trial: overview and study design. Radiology,

258(1):243–253.

Tenerani, M., Imbriani, M., Lizzi, F., Pallotta, S., Quat-

trocchi, M., Retico, A., Saponaro, S., Scapicchio, C.,

Talamonti, C., Tofani, A., Zafaranchi, A., and Fan-

tacci, M. (2024). Sc10.04 evaluation of repeatabil-

ity and robustness of ct-derived radiomic features us-

ing a custom phantom. Physica Medica, 125:103472.

Abstracts of the 5th European Congress of Medical

Physics.

Thakur, S. K., Singh, D. P., and Choudhary, J. (2020). Lung

cancer identiﬁcation: a review on detection and classi-

ﬁcation. Cancer and Metastasis Reviews, 39(3):989–

998.

Tim

on, S., Rinc

on, M., and Mart

ınez-Tom

as, R. (2017). Ex-

tending xnat platform with an incremental semantic

framework. Frontiers in neuroinformatics, 11:57.

Traverso, A., Wee, L., Dekker, A., and Gillies, R. (2018).

Repeatability and reproducibility of radiomic fea-

tures: a systematic review. International Journal of

Radiation Oncology* Biology* Physics, 102(4):1143–

1158.

Van Griethuysen, J. J., Fedorov, A., Parmar, C., Hosny, A.,

Aucoin, N., Narayan, V., Beets-Tan, R. G., Fillion-

Robin, J.-C., Pieper, S., and Aerts, H. J. (2017).

Computational radiomics system to decode the radio-

graphic phenotype. Cancer research, 77(21):e104–

e107.

Vonder, M., Dorrius, M. D., and Vliegenthart, R. (2021).

Latest ct technologies in lung cancer screening: proto-

cols and radiation dose reduction. Translational lung

cancer research, 10(2):1154.

Zhao, Y. R., Xie, X., De Koning, H. J., Mali, W. P., Vliegen-

thart, R., and Oudkerk, M. (2011). Nelson lung cancer

screening study. Cancer Imaging, 11(1A):S79.

Zwanenburg, A., Leger, S., Valli

eres, M., and Lock, S.

Image biomarker standardisation initiative. arXiv

preprint arXiv:1612.07003.

BIOIMAGING 2025 - 12th International Conference on Bioimaging

412