Assessing radiomics feature stability with simulated CT acquisitions Kyriakos Flouris1 Oscar Jimenez-del-Toro2 Christoph Aberle3 Michael Bach3 Roger

2025-05-02 0 0 2.1MB 12 页 10玖币

侵权投诉

Assessing radiomics feature stability with simulated

CT acquisitions

Kyriakos Flouris1,*, Oscar Jimenez-del-Toro2, Christoph Aberle3, Michael Bach3, Roger

Schaer4, Markus Obmann3, Bram Stieltjes3, Henning M ¨uller2,4, Adrien Depeursinge2,5,

and Ender Konukoglu1

1Computer Vision Lab, ETH Zurich, Zurich, Switzerland

2University of Applied Sciences Western Switzerland (HES-SO) Valais, Sierre, Switzerland

3Department of Radiology, University Hospital Basel, University of Basel, Basel, Switzerland

4Faculty of Medicine, University of Geneva (UNIGE), Geneva, Switzerland

5Department of Nuclear Medicine and Molecular Imaging, Lausanne University Hospital, Lausanne, Switzerland

*kﬂouris@vision.ee.ethz.ch

ABSTRACT

Medical imaging quantitative features had once disputable usefulness in clinical studies. Nowadays, advancements in analysis

techniques, for instance through machine learning, have enabled quantitative features to be progressively useful in diagnosis

and research. Tissue characterisation is improved via the “radiomics” features, whose extraction can be automated. Despite the

advances, stability of quantitative features remains an important open problem. As features can be highly sensitive to variations

of acquisition details, it is not trivial to quantify stability and efﬁciently select stable features. In this work, we develop and validate

a Computed Tomography (CT) simulator environment based on the publicly available ASTRA toolbox (www.astra-toolbox.com).

We show that the variability, stability and discriminative power of the radiomics features extracted from the virtual phantom

images generated by the simulator are similar to those observed in a tandem phantom study. Additionally, we show that the

variability is matched between a multi-center phantom study and simulated results. Consequently, we demonstrate that the

simulator can be utilised to assess radiomics features’ stability and discriminative power.

1 Introduction

Computerized quantitative analysis of medical images is emerging as a promising approach in radiological practice and

healthcare research

1–4

. These methods extract measurements quantifying various aspects of the image that include basic

intensity statistics as well as more complicated metrics quantifying spatial intensity heterogeneity. Extracted measurements

are then used as image biomarkers in predicting relevant outcomes. In recent years, numerous researchers demonstrated the

capability of this approach for diagnosis, stratiﬁcation, and prognosis

5,6

. Moreover, since the extraction of measurements as

well as the prediction stage are all algorithmic, quantitative analysis is an efﬁcient approach that can complement radiologists’

visual interpretation and analysis.

Advanced artiﬁcial intelligence techniques

, such as deep learning, take the quantitative analysis approach one step

further

8,9

. They remove the need to engineer measurements to extract from images for a given task. Instead, they optimize

their parameters to extract task-optimal measurements and predict based on them. In the respective language, quantitative

measurements are called “features”. While the optimization requires large number of data samples, i.e., training samples, if

such large datasets exist, deep learning algorithms can provide substantial accuracy gains10.

An important limitation of the quantitative analysis approach is its sensitivity to variations in scanning conditions

. While

the methods aim to extract measurements characterising the underlying tissue composition and microstructure, they are indeed

measurements taken from the image, which is merely a representation of the tissue. Critically, image characteristics heavily

rely on the acquisition details, e.g., resolution, radiation dose, noise, reconstruction algorithm. Depending on the properties

of the algorithm and the measurement, the extracted quantities can be highly sensitive to variations in the image acquisition

parameters

12–14

. This sensitivity inhibits the generalisation capabilities of such measurements. If acquisition details are not

perfectly matched, two different images, even of the same tissue, will yield different measurements. A number of studies have

reported the impact on CT radiomics analysis caused by the variability of acquisition parameters and post-process variables

15–18

Any algorithm or analysis based on these measurements will therefore not be reliable for use with unseen scanners.

The ideal way to study the sensitivity of measurements is to perform test-retest studies

. This would comprise of imaging

a group of subjects imaged under different acquisition details. To study sensitivity of a measurement, values extracted from

arXiv:2210.02759v1 [physics.comp-ph] 6 Oct 2022

corresponding images would be compared. When new measurements or new algorithms to extract measurements are proposed,

they would be studied the same way. As this is not feasible for various imaging modalities, such as Computed Tomography

(CT) due to the radiation exposure of patients in these studies, anthropomorphic printed phantoms have been proposed for CT

variability studies20–22.

Phantom studies have been successfully used for various imaging modalities. Especially for CT, advances in 3D printing

technologies allow printing volumetric patient images using materials with attenuation properties comparable to human tissue.

Recent work reported variability studies using such phantoms23–25.

While phantoms make it possible to study variability without imaging cohorts, they still require acquiring and imaging

phantoms. This can be costly as well as resource and time consuming. In this work, we study whether sensitivity analysis using

advanced in-silico CT simulators can yield similar results to real phantom imaging studies. To this end, a CT-scan simulator

environment was set up using the publicly available

26,27

ASTRA toolbox (www.astra-toolbox.com). Using a high-dose

CT-image as input, the simulator outputs raw projections, which can be manipulated accordingly. For example, stationary and

uncorrelated noise can be added. Additionally, the simulator allows for some freedom in geometrical parameters such as the

number of projections, slice thickness, and distances. The CT-image can be reconstructed with a variety of algorithm choices,

e.g. ﬁltered back-projection and simultaneous iterative reconstruction technique.

The method is compared with an empirical anthropomorphic phantom variability study published in

. In this unique setup,

the simulated phantom study is performed using the same original image from which the anthropomorphic phantom was printed

and the study in

conducted. In a sense, this can be viewed as the theoretical replication. The simulator environment was

implemented to reconstruct images at different noise levels, reconstruction algorithms, and number of projections. To mimic

repetition and introduce variability, each simulation parameter set was repeated via a variation of the Poisson noise random

seed. For the simulated images, radiomics features were extracted and analysed. As the same source image is used for both the

empirical phantom study and this work, direct comparison of the results of sensitivity analyses is possible.

The next section describes the CT simulator environment method including a brief introduction of the anthropomorphic

phantom and the phantom study. In the results section a comprehensive validation and comparison of the simulator with respect

to the phantom study is presented. Furthermore, a stability and discriminative power analysis and discussions can be found in

the same section. The paper is summarised in the conclusions section.

2 Method

First, we introduce the details of the novel anthropomorphic phantom created for the tandem phantom study

. A high dose

CT-scan of this phantom is used as the simulator input. Second, the extracted radiomics features, the principal component

analysis and the simulator environment are described in detail.

2.1 Anthropomorphic phantom and phantom CT acquisitions

Here, we provide brief details of the anthropomorphic phantom study presented in

for completeness. For further details, we

refer the reader to the original publication.

A realistic radio-opaque three-dimensional phantom was designed from real patient CT data. Namely, the compilation

of a half-mirrored lung including a tumor and an abdominal liver section with a metastasis from a colon carcinoma

. The

phantom was manufactured via stacking sheets of printed aqueous potassium iodide solution on paper

. The lung tumor section

is a replication of a publicly available patient data set for radiomics phantoms, from the Image Biomarker Standardization

Initiative

. The lung section was neither used in this work nor the tandem phantom CT study. Tissue equivalent attenuation at

a deﬁned energy spectrum was calibrated at 120 kVp. The contrast resolution of the printing technique in the phantom goes

from -100 to 1000 Hounsﬁeld units (HU). Overall, no structures can be represented whose HU is below this paper-induced

threshold. To test the contrast resolution, a circular intensity ramp was printed in the phantom running through an HU range of

0 to 1000. A reliable resolution of 2 HU difference was achieved. Consequently the abdominal region was adequately depicted

for a quantitative analysis within the printed HU range.

The phantom was imaged with a Siemens SOMATOM Deﬁnition Edge CT scanner (SSDE). To deﬁne the acquisition and

image reconstruction parameters, a survey of clinical CT protocols was performed including 9 radiological institutes. All the

CT scans in that study were acquired with the same acquisition parameters, which resulted in an approximate CT dose index of

10mGy. Namely, a tube voltage of 120 kVp, a helical pitch factor of 1.0, a 0.5 second rotation time, and a tube current time

product of 147 mAs. No automatic tube current modulation was used.

Typical reconstruction parameter settings for clinical protocols in thoracic and abdominal oncology were varied for

the phantom study as follows: Reconstruction algorithm, iterative reconstruction (IR) or ﬁltered back projection (FBP);

reconstruction kernel, 2 standard soft tissue kernels per algorithm; slice thickness in millimeter, 1, 1.5, 2, 3; and slice spacing

in millimeter, 0.75, 1, and 2. Series reconstructed with an IR algorithm used an ADMIRE (advanced modeled iterative

reconstruction) at strength level 3. In total, 8 groups of parameter variations were selected for the phantom study to assess

2/12

their impact on classic radiomics features. Initially, 20 repetition scans were performed without re-positioning of the phantom,

followed by 10 repetitions with re-positioning between each measurement. Therefore, 30 distinct acquisitions were performed

for each of 8 parameter variation groups.

In the abdominal section six 3D regions of interest (ROIs) were manually annotated by a board-certiﬁed radiologist using

a thin-section phantom series with 2 mm slice thickness and 1 mm spacing. The ROIs were annotated conservatively, well

within the margins, thus no cross-check step of the annotations was performed by other radiologists. A polygonal outline was

used on all slices individually to deﬁne the ROIs. The six ROI binary masks were stored in a 3D volume NIfTI format. Two

normal liver tissue regions, two cysts, a hemangioma, and a liver metastasis from a colon carcinoma were included during the

annotation process, regions can be found in Figure 1. Further details of the annotated regions and the 8 variation groups can be

found in Jimenez-del-Toro et al.23.

hemangioma

metastasis

normal_2

normal_1

cyst_2

cyst_1

Figure 1. Annotated regions of interest on the anthropomorphic phantom.

A multi-center phantom CT study was also carried out with 13 different scanners at selected locations in Switzerland.

The scanners used were two Siemens SOMATOM Deﬁnition Edge, two Siemens SOMATOM Deﬁnition Flash, a Siemens

SOMATOM Edge Plus, a Siemens SOMATOM X.Cite, a Philips Brilliance iCT, a GE BrightSpeed S, a Philips Brilliance CT

64, a GE Revolution Evo, a GE Revolution Apex, a Canon Aquilion Prime SP and a Canon Aquilion CXL. The same protocol

was implemented (as closely as possible) in all acquisitions. A tube voltage of 120 kVp, a helical pitch factor of 1.0 and a 0.5

second rotation time were used. The tube current time product was adjusted accordingly to achieve the required dose of 10mGy.

The IR reconstruction algorithm was used with slice thicknesses 2 or 2.5 millimeter and slice spacing 1 or 1.25 millimeter.

2.2 Radiomics feature extraction and principal component analysis

From both the simulated and phantom CT scans, a total of 86 radiomics features were extracted in 3D from the manually

segmented ROIs using the open source Pyradiomics python toolkit

. Deﬁnitions for the radiomics features are available in

the Pyradiomics documentation online (https://pyradiomics.readthedocs.io/en/latest/features.html). The 86 features extracted

include 18 ﬁrst-order statistics, 22 grey level co-occurrence matrices, 14 grey level dependence matrices, 16 grey level run

length matrices and 16 grey level size zone matrices, as described brieﬂy in the Appendix B. Radiomics features parameters

were set to their default values. More speciﬁcally, no ﬁlter was applied to the input image and a ﬁxed bin width of 25 was

used for the discretisation of the image grey level. Fixed bin size discretisation is deﬁned such that a new bin is assigned for

every intensity interval within the bin width starting at the lowest occurring intensity. Additionally, no normalization, no spatial

resampling, no resegmentation were performed and no HU cutoffs were used within the ROIs for the extraction. The distance

between the center voxel and the neighbor, for which angles should be generated, was set to one pixel. Furthermore, for the ﬁrst

order radiomics the voxel array shift parameter was set to zero, for the grey level co-occurrence matrices the co-occurrences

was assessed in two directions per angle, which results in a symmetrical matrix and for the grey level dependence matrices no

cutoff value for dependence was set, i.e. a neighbouring voxel was always considered independent.

For the phantom CT acquisitions, an analysis was carried out via the principal component analysis (PCA). The ﬁrst two

principal components of the 86 radiomics features from all 240 phantom CT acquisitions are shown in Figures 6and 7with

black markers. The ROIs can be separated into 4 distinct tissue classes, i.e. normal liver tissue, cyst, hemangioma, and liver

metastasis. The differences between the four ROI classes (inter-class variation) are larger than all CT parameter variations

(intra-class variation). ROIs from the normal liver tissue class are closer in the feature space than those from the other classes.

All four classes remain linearly separable despite the CT parameter variations.

Furthermore, the Wilcoxon statistic

was used to assess the stability and discriminative power of isolated radiomics

features

. We set a threshold of

< 1 to indicate a stable comparison. The top 10 ranked features of the phantom CT

acquisitions are shown on the right-hand of the appendix Table 3.

3/12

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AssessingradiomicsfeaturestabilitywithsimulatedCTacquisitionsKyriakosFlouris1,*,OscarJimenez-del-Toro2,ChristophAberle3,MichaelBach3,RogerSchaer4,MarkusObmann3,BramStieltjes3,HenningM¨uller2,4,AdrienDepeursinge2,5,andEnderKonukoglu11ComputerVisionLab,ETHZurich,Zurich,Switzerland2UniversityofAppliedS...

展开>> 收起<<

Assessing radiomics feature stability with simulated CT acquisitions Kyriakos Flouris1 Oscar Jimenez-del-Toro2 Christoph Aberle3 Michael Bach3 Roger.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Assessing radiomics feature stability with simulated CT acquisitions Kyriakos Flouris1 Oscar Jimenez-del-Toro2 Christoph Aberle3 Michael Bach3 Roger

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: