A novel lung radiomics feature for characterizing resting heart rate and COPD stage evolution based on radiomics feature combination strategy

1.   Introduction

The resting heart rate (HR, beats per minute) variability is an important marker of the heart's neurophysiologic condition ^[1]. The resting HR can be measured to reflect the level of cardiopulmonary. Lung disease is a complex and diverse disease. As a common and non-infectious lung disease, chronic obstructive pulmonary disease (COPD) presents a preventable, treatable and progressive chronic disease with debilitating lung conditions characterized by persistent airflow limitation ^[2,3]. Due to the persistent airflow limitation of COPD, autonomic regulation of resting HR can be influenced ^[4]. Compared with people without COPD, patients with COPD cannot get enough air, which may increase the resting HR to get enough oxygen supply. As a result, the resting HR is an upward trend with the development of COPD severity ^[5]. After using a bronchodilator, the pulmonary function test (PFT) can assess COPD severity. The assessment parameters in PFT are the forced expiratory volume in 1 second/forced vital capacity (FEV1/FVC), and FEV1 % predicted ^[6]. The HR variability in COPD patients has been deeply studied ^[1,5,7,8]. There is no correlation between resting HR variability and FEV1 in COPD patients ^[1]. However, the COPD stage is determined by FEV1/FVC, and FEV1% predicted, and the relationship between the resting HR variability and the COPD stage is not revealed. Because the resting HR is an upward trend with the development of COPD severity, the resting HR can improve the risk prediction in COPD patients beyond that of PFT alone ^[5]. The HR is analyzed in the process of the Chester step tests to determine reliability and correlation with PFT results in COPD patients, according to advanced stages of the Chester step test and the number of steps ^[7], respectively. The effect of pulmonary rehabilitation on HR variability at peak exercise is revealed in COPD patients ^[8]. Symbolic analysis and complexity index of the HR variability are analyzed to assess cardiac autonomic modulation in COPD patients ^[9]. Association between the predictors of functional capacity and HR off-kinetics is also studied in COPD patients ^[10]. The relationship of the HR variability and the severity of COPD in PiZ alpha1-antitrypsin deficiency has also been revealed ^[11]. As the gold standard, PFT can only diagnose and evaluate COPD ^[6], but it cannot get the anatomical structure of the lung region. PFT may cause missed diagnosis of early-stage COPD or overdiagnosis in primary care ^[12]. Compared with PFT, chest computed tomography (CT) images can provide more information, such as the specific anatomical structure of the lung, the location, and the morphology of diseases. Chest CT images can provide the lung anatomical structure, including the trachea, the blood vessels, pulmonary lobes, and lung texture information. The lung anatomical structure and texture information can be quantitatively calculated from chest CT images to further analyze COPD ^[13,14,15]. Therefore, CT has been regarded as the most effective modality for characterizing and quantifying COPD ^[16].

Lung radiomics features calculated from chest CT images have been used in the spirometric assessment of emphysema presence and COPD severity ^[17]. The emerging role of radiomics in COPD has also been proposed ^[18]. The COPD radiomics features can provide feature classes about COPD extracted from original and derived images. Although the HR variability in COPD ^{[1,5,7,8,9,10]} has been fully studied, the relationship between the resting HR variability and COPD radiomics features remains unclear. Our contributions in this paper are briefly described as follows:

• The relationship between the COPD radiomics features and the resting HR is revealed.

• A new COPD radiomics feature combination algorithm is proposed to improve the significance among different COPD stages.

• A novel lung radiomics features with the dominant selected COPD radiomics features characterize both the resting heart rate and the COPD stage evolution.

2.   Materials and methods

This Section describes the Materials in Section 2.1 and methods in Section 2.2 (Figures 1−4, Eqs (1)−(3) and Table 1).

Figure 1.  Subjects selection flow diagram and the number of subjects at different COPD stages. Figure 1(A): subjects selection flow diagram, Figure 1(B): number of subjects at different COPD stages, and Figure 1(C): the changing trend of resting HR with COPD stage evolution.

DownLoad: Full-Size Img PowerPoint

Figure 2.  Overall block diagram of this study. Figure 2(A) shows the lung region segmentation, Figure 2(B) shows the calculation of the radiomics feature, and Figure 2(C) shows the selection, features combination, and statistical analysis of COPD radiomics features, respectively.

DownLoad: Full-Size Img PowerPoint

Figure 3.  Typical lung region segmentation results from the original chest HRCT images in the coronal, transverse, and sagittal planes. Figure 3(A)−(C) shows the original chest HRCT images in the coronal, transverse, and sagittal planes. Figure 3(D)−(F) shows the lung region segmentation results in the corresponding plane. The red mask is the right lung parenchyma mask, and the green one is the left.

DownLoad: Full-Size Img PowerPoint

Figure 4.  The detailed process of the COPD radiomics features calculation. The COPD radiomics features are calculated by the predetermined classes of radiomics features and the two types of lung region images.

DownLoad: Full-Size Img PowerPoint

Table 1.  The algorithm of the COPD radiomics features combination strategy.

	Detailed implementation process
Input	Coefficient array [β1, β2, …., βi] Initialization: preset significant condition (P = 0.005)
Output	Lung radiomics feature ${F_{lung{\text{ }}radiomics}}$; N
Computation	Absolute value operation for the coefficient array [|β1|, |β2|, …., |βi|] Sorting operation for the coefficient array [max {|β1|, |β2|, …., |βi|}, …., min {|β1|, |β2|, …., |βi|}] for N = 2; N ≤ 13; N ← N + 1 do       {             Form a new coefficient array             $\overrightarrow {{\beta _N}} = \underbrace {[max\{ |{\beta _1}\left| {, {\text{ }}} \right|{\beta _2}\left| {, {\text{ }} \ldots., {\text{ }}} \right|{\beta _i}\left| {\}, \ldots., min\{ } \right|{\beta _1}\left| {, {\text{ }}} \right|{\beta _2}\left| {, {\text{ }} \ldots., {\text{ }}} \right|{\beta _i}|\} ]}_N$             Construct the candidate lung radiomics feature ${Z}_{N}={\displaystyle \sum _{i=1}^{N}{\overrightarrow{\beta }}_{N}·{x}_{i}}$ (4)             Calculate the significances of ${Z_N}$ among different COPD stages             if all significances < (P = 0.05)                   ${F_{lung{\text{ }}radiomics}} = {Z_N}$                   break from for             end if       } end for return {${F_{lung{\text{ }}radiomics}}$; N }

---

 | Show Table

DownLoad: CSV

2.1.   Materials

Figure 1 shows the Chinese subjects selection flow diagram, the number of subjects at different COPD stages, and the changing trend of resting HR with COPD stage evolution. 231 Chinese subjects aged 40−79 are included in this study. The 231 subjects who rigorously followed the inclusion and exclusion criteria were enrolled in the national clinical research center of respiratory diseases ^[21]. The 231 subjects with full inspiration underwent chest high-resolution CT (HRCT) scans (manufacturer: TOSHIBA, kVp:120 kV, X-ray tube current:40 mA, slice thinkness:1.0 mm, window center: −600, window width: 1250) from May 25, 2009, to January 11, 2011, are included in this study. In addition, the 231 subjects, after 15 minutes of rest, underwent 12 Leads of ECG on the same day for the resting HR measurement.

The resting HR of all subjects affected by COPD and the abnormal resting HR caused by other diseases were excluded in our study. Diagnosis of COPD classification was from stage I to IV according to Global Initiative for Chronic Obstructive Lung Disease (GOLD) 2008 criteria accepted by the American Thoracic Society and the European Respiratory Society. According to our definition, at risk of COPD (stage 0) was diagnosed. The definition for stage 0 is that chronic cough and sputum production are for at least three months in each of two consecutive years without any other condition explaining the cough and a post-bronchodilator FEV1/FVC ≥ 0.7 and FEV1 ≥ 80% predicted ^[21]. The resting HR of the 231 subjects need further screening to exclude the abnormal resting HR outside the interval ([60,100], beats per minute) ^[5]. After excluding the abnormal resting HR, 196 subjects are used to determine a lung radiomics feature for characterizing resting heart rate and COPD stage evolution in this study. COPD stage 0−IV has 50, 46, 58, 32 and 10 subjects, respectively.

The ethics committee had approved this study of the national clinical research center of respiratory diseases in Guangzhou medical university, China. Each subject had been provided written informed consent by the first affiliated hospital of Guangzhou medical university before chest HRCT scans and 12 Leads of ECG.

2.2.   Methods

This Section describes the methods including the lung region segmentation (Section 2.2.1), COPD radiomics features calculation (Section 2.2.2), COPD radiomics features selection (Section 2.2.3), and COPD radiomics features combination strategy (Section 2.2.4) in detail.

Figure 2 shows the overall block diagram of this study. Figure 2(A) shows that the trained segmentation model segments the lung parenchyma mask images (512 × 512 × N) from the original chest HRCT images (512 × 512 × N). Figure 2(B) shows the calculation of the radiomics features based on the lung parenchyma images. The original lung parenchyma images are obtained based on the original chest HRCT images and lung parenchyma mask images. Then, the original lung parenchyma images are filtered to get the derived lung parenchyma images. The original chest HRCT images and the derived lung parenchyma images are used to calculate the COPD radiomics features according to a predetermined class of radiomics features. Figure 2(C) shows the selection of lung radiomics features related to HR and the selected COPD radiomics features combination to obtain lung radiomics features that characterize HR and COPD stage evolution. The COPD radiomics features and the corresponding resting HR data of 196 subjects are normalized together before importing the Lasso model. 13 selected COPD radiomics features are generated by the Lasso model. Finally, a COPD radiomics features feature is constructed by the proposed radiomics combination strategy for fully characterizing the resting HR and COPD stage evolution.

2.2.1.   Lung region segmentation

The lung region needs to be accurately segmented from the chest HRCT images to calculate the COPD radiomics features. U-Net has been widely used to segment biomedical images ^[22,23,24]. Some new convolution networks also are proposed, such as a PedNet for image segmentation ^[25]. Similarly, some networks have made innovations in the application, such as residual networks for the image quality assessment ^[26]. Based on U-Net and residual networks, a fully automatic segmentation model ^[27] named U-net (R231) is used to segment the lung region from the chest HRCT images in Figure 2(A). The U-net (R231), which had been trained by human chest CT images, is a U network model with residual building block (ResU-Net) ^[28]. The architecture of the ResU-Net model is described in detail in our previous study ^[28]. The trained ResU-Net is available on the website https://github.com/JoHof/lungmask.

Figure 3 shows the typical lung region segmentation results from the original chest HRCT images in the coronal,transverse,and sagittal planes. The images without red and green color are the original chest HRCT images,and the images with red and green colors are the corresponding lung region segmentation results. All slices of lung region segmentation results have been carefully checked and modified by three experienced radiologists using a tool named ITK-SNAP. The ITK-SNAP can be available on the website,http://www.itksnap.org/pmwiki/pmwiki.php?n=Downloads.SNAP3. The three experienced radiologists consider that all lung region segmentation results are acceptable for calculating COPD radiomics features.

2.2.2.   COPD radiomics features calculation

COPD radiomics features are calculated based on the original and derived lung parenchyma images. Therefore, the lung parenchyma images should be extracted from the chest HRCT images. The method of extracting original lung parenchyma images is based on the chest HRCT images and can refer to our previous study ^[28,29].

Wavelet provides the different scales of the chest images for the image analysis by using its multi-resolution decomposition ^[30,31]. Laplacian of Gaussian filter (LoG) ^[32,33] as an edge enhancement filter can emphasize areas of gray level change in images which is crucial for COPD chest images. Because of their advantages, the wavelet filter and LoG filter are considered for generating the derived lung parenchyma images in this study.

Figure 4 shows the detailed process of the COPD radiomics features calculation features. First,the wavelet and LoG filters are applied to the original lung parenchyma images. Specifically,the wavelet filter yields 8 (2³) decompositions per level with all possible combinations of applying either a high pass filter (H) or low pass filter (L) in each of the three dimensions. Therefore,the wavelet filter can derive eight types of derived lung parenchyma images. The LoG filter with sigma 1.0 to 5.0. can derive five types of derived lung parenchyma images. Finally,the COPD radiomics features are calculated based on the original and the 13 derived lung parenchyma images shown in Figure 3. Specifically,PyRadiomics (version 3.0.,a radiomics calculation model) ^[34] is applied to calculate the COPD radiomics features. The PyRadiomics is available on the website https://pyradiomics.readthedocs.io/en/latest/index.html, and the website also has given detailed explanations of radiomics. Finally, 1316 COPD radiomics features of each subject are obtained.

2.2.3.   COPD radiomics features selection

This paper uses the Lasso model ^[35,36] to select the COPD radiomics features related to HR. The mathematical form of the Lasso model is shown in expression (1):

where $x_{ij}^*$ is the value of the independent variable after normalization, $y_i^*$ is the value of the dependent variable, λ is a penalty parameter (λ ≥ 0), and β_j is the regression coefficient vector, i∈[1, n], and j∈[0, p].

A standard R package "lars 1.2" (parameter: type = "lasso", and use.Gram = FALSE) is performed by an operating environment RStudio to select the independent variable. A tenfold cross-validation (a standard R package "cv. Lars" with parameter: type = "lar", K = 10, and use.Gram = FALSE) is used to select the fraction (the minimus cross-validated MSE). The dependent variable is the resting HR∈[60,100], and the independent variable is the COPD radiomics features. The size of COPD radiomics features is 196 × 1316 (196 subjects and 1316 COPD radiomics features of each subject). However, the resting HR and the COPD radiomics features should be normalized before importing the Lasso model. The mathematical form of the normalization is shown in Eq (2):

where ${x_{ij}}$is the independent variable (COPD radiomics features) before normalization, ${y_i}$ is the dependent variable (the resting HR) before normalization, $\overline {{x_j}}$is the mean of the independent variable ${x_{ij}}$, ${x_{jmax}}$ is the maximum of the independent variable ${x_{ij}}$, ${x_{jmin}}$ is the minimum of the independent variable ${x_{ij}}$, $\overline {{y_i}}$is the mean of the dependent variable ${y_i}$, ${y_{imax}}$ is the maximum of the dependent variable ${y_i}$, and ${y_{imin}}$ is the minimum of the dependent variable ${y_i}$.

2.2.4.   COPD radiomics features combination strategy

The COPD stages Ⅲ and Ⅳ are taken as one stage (stage Ⅲ & Ⅳ) in this paper to balance the data at different COPD stages and meet the statistical need. The lung radiomics feature Y_k is constructed by the following combination Eq (3) to observe the significant change among the different types of lung parenchyma images.

where k is the type of lung parenchyma images, N is the number of the selected COPD radiomics features belonging to one type of lung parenchyma images, and β_i is the coefficient of the selected radiomics x_i generated by the Lasso model. K = 1 denotes all types of lung parenchyma images. K = 2 denotes the original lung parenchyma images. K = 3 and 4 denote the derived lung parenchyma images generated from the LoG and wavelet filters.

A COPD radiomics features combination strategy is proposed in this paper to improve the significance among the COPD stages. Table 1 shows the specific algorithm of the COPD radiomics features combination strategy. The idea of the algorithm is to use the least COPD radiomics features to reflect the significances among different COPD stages and retain COPD radiomics features with the maximum coefficient array, which can reflect the resting HR.

Specifically, the coefficient array is the coefficients generated from the Lasso model. After initializing the preset significant condition, each coefficient in the coefficient array is changed to an absolute value. The coefficient array [β₁, β₂, …., β_i] turns to [|β₁|, |β₂|, …., |β_i|]. The absolute values are sorted from largest to smallest, getting the pending coefficient array [max {|β₁|, |β₂|, …., |β_i|}, …., min {|β₁|, |β₂|, …., |β_i|}]. Next, the selected number N is initialized to 2, which means that the first two coefficients are chosen in the pending coefficient array. The N-selected coefficients from the pending coefficient array form a new coefficient array $\overrightarrow {{\beta _N}}$. Then a candidate lung radiomics feature ${Z_N}$ is constructed by the new coefficient array $\overrightarrow {{\beta _N}}$ and its COPD radiomics features using the linear combination. The linear combination Eq (4) is shown in Table 1. In the linear combination Eq (4) ${x_i}$ is the selected COPD radiomics features. Finally, the candidate lung radiomics feature's significances among different COPD stages are calculated. Dunn's multiple comparisons test in the statistical software GraphPadPrism (8.0.1) calculates all significances among different COPD stages. If all significances among different COPD stages < the preset significant condition (P = 0.05), the candidate lung radiomics feature ${Z_N}$ is considered as the lung radiomics feature, which characters the resting HR and COPD stage evolution.

3.   Results

All the significances of the 13 selected COPD radiomics features, the constructed lung radiomics features Y1−Y4, and the constructed lung radiomics features F1 and F2 among different COPD stages are analyzed in this Section.

3.1.   The selected COPD radiomics features based on Lasso

Figure 5 shows the cross-validated mean square error (MES) with the fraction of final L1 norm by standard R package "cv. Lars" (K = 10, tenfold cross-validation). The optimal fraction (0.0909) is determined when the cross-validated MSE takes the minimum value. After the tenfold cross-validation, 13 COPD radiomics features of each subject are selected from the 1316 COPD radiomics based on Lasso.

Figure 5.  The cross-validated MES with the fraction of final L1 norm based on the Lasso.

DownLoad: Full-Size Img PowerPoint

Table 2 shows the name, class, definition and coefficient of the 13 selected COPD radiomics features. We rename the 13 selected COPD radiomics features as Radiomics 1−13 for convenient description in this paper. The coefficients of the 13 selected COPD radiomics features are generated from the Lasso model. Radiomics 1−3 are the radiomics of the original lung parenchyma images, and radiomics 4−8 are the radiomics of the derived lung parenchyma images generated by the LoG filter with the sigma 1.0−5.0. Radiomics 9−13 are the radiomics of the derived lung parenchyma images generated by the wavelet filter.

Table 2.  Name, class, coefficient and definition of the 13 selected COPD radiomics features.

Name of the selected COPD radiomics features	Class	Coefficient	Definition
original_shape_Elongation	Shape Features	0.0502	Radiomics1
original_shape_MajorAxisLength	Shape Features	−0.0128	Radiomics2
original_shape_Maximum2DDiameterColumn	Shape Features	−0.0715	Radiomics3
log.sigma.2.0.mm.3D_firstorder_Kurtosis	First Order Features	−0.0024	Radiomics4
log.sigma.2.0.mm.3D_firstorder_Maximum	First Order Features	−0.0038	Radiomics5
log.sigma.2.0.mm.3D_glszm_GrayLevelVariance	GLSZM 1 Features	−0.0231	Radiomics6
log.sigma.4.0.mm.3D_glszm_GrayLevelVariance	GLSZM 1 Features	−0.0132	Radiomics7
log.sigma.5.0.mm.3D_glszm_SizeZoneNonUniformity	GLSZM 1 Features	−0.0149	Radiomics8
wavelet.HLH_firstorder_Skewness	First Order Features	0.0069	Radiomics9
wavelet.HLH_glcm_MCC	GLCM 2 Features	0.0094	Radiomics10
wavelet.HLH_glrlm_LongRunLowGrayLevelEmphasis	GLRLM 3 Features	−0.0156	Radiomics11
wavelet.HLH_glszm_LargeAreaLowGrayLevelEmphasis	GLSZM 1 Features	−0.0590	Radiomics12
wavelet.LLL_firstorder_10Percentile	First Order Features	−0.2494	Radiomics13
*Note: 1 Gray Level Size Zone Matrix. 2 Gray Level Cooccurrence Matrix. 3 Gray Level Run Length Matrix.

---

 | Show Table

DownLoad: CSV

Table 2 gives the relationship (coefficient) between the 13 selected COPD radiomics features and resting HR. The symbol "-" indicates a positive correlation of the selected COPD radiomics feature and resting HR, and the omitted symbol "+" indicates a positive correlation. Since all COPD radiomics features have been normalized before the selection, the coefficient can represent the importance of its corresponding COPD radiomics features. Figure 6 shows the coefficient of the 13 selected COPD radiomics features and the number of the selected COPD radiomics features in each class. Figure 6(A) shows the order of importance is Radiomics13, 3, 12, 1, 6, 11, 8, 7, 2, 10, 9, 5, 4. Therefore, Radiomics13 is the dominant COPD radiomics feature to affect the resting HR. Figure 6(A) shows the number of the selected COPD radiomics features in each class. Although the numbers of the first order and GLSZM features are the same, the dominant COPD radiomics feature Radiomics13 belongs to the first order features. Therefore, the first order features have a greater impact on the resting HR than GLSZM features.

Figure 6.  The picture of the coefficient of the 13 selected COPD radiomics features and the number of the selected COPD radiomics features in each class. Figure 6(A): the coefficient of the 13 selected COPD radiomics features, and Figure 6(B): the number of the selected COPD radiomics features in each class.

DownLoad: Full-Size Img PowerPoint

Figure 7(A)−(M) shows the changing trend with a boxplot of the 13 selected COPD radiomics features among different COPD stages. Table 3 shows significant differences in the selected COPD radiomics features among some different COPD stages, including Radiomics1, 3, 4, 6, 8, 9, 11, 12 and 13. However, there are no significant differences among all COPD stages in Radiomics2, 5, 7 and 10. Therefore, the results of Radiomics1, 3, 4, 6, 8, 9, 11, 12 and 13 are further analyzed.

Figure 7.  The changing trend of the 13 selected COPD radiomics features among different COPD stages.

DownLoad: Full-Size Img PowerPoint

Table 3.  Using Dunn's multiple comparisons test, the adjusted P-value of the13 selected COPD radiomics features among different COPD stages.

COPD stage comparison	Radiomics1	Radiomics2	Radiomics3	Radiomics4	Radiomics5
Stage 0 vs. Stage Ⅰ	0.0067	0.0675(ns)	< 0.0001	0.0018	> 0.9999 (ns)
Stage 0 vs. Stage Ⅱ	0.0034	0.0951(ns)	0.0001	0.0062	> 0.9999 (ns)
Stage 0 vs. Stage Ⅲ & Ⅳ	< 0.0001	> 0.9999 (ns)	0.0003	> 0.9999 (ns)	> 0.9999 (ns)
Stage Ⅰ vs. Stage Ⅱ	> 0.9999 (ns)	> 0.9999 (ns)	> 0.9999 (ns)	> 0.9999 (ns)	> 0.9999 (ns)
Stage Ⅰ vs. Stage Ⅲ & Ⅳ	0.0495	0.0700(ns)	> 0.9999 (ns)	0.0425	> 0.9999 (ns)
Stage Ⅱ vs. Stage Ⅲ & Ⅳ	0.03318	0.0987(ns)	> 0.9999 (ns)	0.1213(ns)	> 0.9999 (ns)
COPD stage comparison	Radiomics6	Radiomics7	Radiomics8	Radiomics9	Radiomics10
Stage 0 vs. Stage Ⅰ	> 0.9999 (ns)	> 0.9999 (ns)	0.0295	> 0.9999 (ns)	> 0.9999 (ns)
Stage 0 vs. Stage Ⅱ	0.06115 (ns)	> 0.9999 (ns)	0.3687 (ns)	> 0.9999 (ns)	> 0.9999 (ns)
Stage 0 vs. Stage Ⅲ & Ⅳ	> 0.9999 (ns)	> 0.9999 (ns)	> 0.9999 (ns)	0.0014	0.5143 (ns)
Stage Ⅰ vs. Stage Ⅱ	0.0508 (ns)	> 0.9999(ns)	> 0.9999 (ns)	> 0.9999 (ns)	> 0.9999 (ns)
Stage Ⅰ vs. Stage Ⅲ & Ⅳ	> 0.9999 (ns)	> 0.9999 (ns)	0.3202 (ns)	0.0070	0.8122 (ns)
Stage Ⅱ vs. Stage Ⅲ & Ⅳ	0.0480	0.8283 (ns)	> 0.9999 (ns)	0.0125	> 0.9999 (ns)
COPD stage comparison	Radiomics11	Radiomics12	Radiomics13
Stage 0 vs. Stage Ⅰ	> 0.9999 (ns)	0.0002	< 0.0001
Stage 0 vs. Stage Ⅱ	> 0.9999 (ns)	0.0081 (ns)	< 0.0001
Stage 0 vs. Stage Ⅲ & Ⅳ	0.0330	< 0.0001	< 0.0001
Stage Ⅰ vs. Stage Ⅱ	> 0.9999(ns)	> 0.9999 (ns)	0.4088 (ns)
Stage Ⅰ vs. Stage Ⅲ & Ⅳ	0.4837(ns)	0.3096 (ns)	< 0.0001
Stage Ⅱ vs. Stage Ⅲ & Ⅳ	0.1819(ns)	0.0089	0.0030
*Note: ns: no significance.

---

 | Show Table

DownLoad: CSV

Figure 7(A) shows that Radiomics 1 increases with the development of COPD stages. The mean ± the standard error of mean (SEM) of Radiomics 1 from the COPD stage 0 to Ⅲ & IV is 0.8068 ± 0.0105, 0.8575 ± 0.0088, 0.8585 ± 0.0081, and 0.8950 ± 0.0084, respectively. However, there is no significant change only between COPD stage I and COPD stage Ⅱ (P > 0.9999). Figure 7(C) shows that Radiomics 3 increases with the COPD stages evolution. Compared to COPD stage 0, Radiomics 3 in COPD stage I, Ⅱ, Ⅲ & IV rises. The mean ± SEM of Radiomics 3 from the COPD stage 0 to Ⅲ & IV is 314.5 ± 3.289,336.1 ± 2.425,332.7 ± 2.580 and 333.5 ± 2.607, respectively. The significant change exists only between COPD stage 0 and COPD stage I, Ⅱ, Ⅲ & IV (P < 0.0001, P = 0.0001, P = 0.0003). Figure 7(D) and Table 3 show that Radiomics 4 significantly increases form COPD stage 0 to COPD stage I, Ⅱ (P = 0.0018 and 0.0062). The mean ± SEM of Radiomics 4 from the COPD stage 0 to Ⅲ & IV is 6.552 ± 0.0781, 6.919 ± 0.0691, 6.884 ± 0.0644 and 6.662 ± 0.0576, respectively. Figure 7(F) shows that the mean of Radiomics 6 reaches the maximum at COPD stage Ⅱ (32.67 ± 0.3414) among all the COPD stages. The mean ± SEM of Radiomics 6 at the COPD stage 0, COPD stage I, and COPD stage Ⅲ & IV is 31.41 ± 0.3988, 31.55 ± 0.4403, 31.42 ± 0.4855, respectively. There are significant changes of Radiomics 6 between COPD stage Ⅱ and Ⅲ & IV (P = 0.0480) shown in Table 3. Figure 7(H) and Table 3 show that Radiomics 8 significantly arises only from COPD stage 0 to COPD stage I (P = 0.0295). The mean ± SEM of Radiomics 8 from the COPD stage 0 to Ⅲ & IV is 6559 ± 235.6, 7466 ± 234.5, 7139 ± 202.7 and 6851 ± 191.4, respectively. Figure 5(J) and Table 3 show that Radiomics 9 significantly arises from COPD stage 0, I, Ⅱ to COPD stage Ⅲ & IV (P = 0.0014, P = 0.0070, P = 0.0125). The mean ± SEM of Radiomics 9 from the COPD stage 0 to Ⅲ & IV is 0.1388 ± 0.0085, 0.1458 ± 0.0074, 0.1512 ± 0.0072 and 0.1875 ± 0.0096, respectively. Figure 7(K) and Table 3 show that Radiomics 11 significantly arises from the COPD stage 0 to Ⅲ & IV (P = 0.0330). The mean ± SEM of Radiomics11 at the COPD stage 0 and Ⅲ & IV is 0.1868 ± 0.0064 and 0.2140 ± 0.0102, respectively. Figure 7(L) and Table 3 show that Radiomics 12 significantly arises from the COPD stage 0 to COPD stage I (P = 0.0002) and from COPD stage 0, Ⅱ to COPD stage Ⅲ & IV (P < 0.0001, P = 0.0089). The mean ± SEM of Radiomics 12 from the COPD stage 0 to Ⅲ & IV is 2.956×107 ± 0.3952 × 107, 6.869 × 107 ± 1.006×107, 4.515 × 107 ± 0.3997×107 and 10.83 × 107 ± 1.797×107, respectively. Figure 5(N) shows that Radiomics 13 decreases with the development of COPD stages. The mean ± SEM of Radiomics 13 from the COPD stage 0 to Ⅲ & IV is 2575 ± 9.006, 2624 ± 6.475, 2662 ± 6.043 and 2718 ± 8.700, respectively. However, there is no significant change between COPD stage I and COPD stage Ⅱ (P = 0.4088).

3.2.   The lung radiomics features based on the proposed combination strategy

Four lung radiomics features Y1−Y4 are constructed using the Eq (3). Specifically, the lung radiomics feature Y1 is constructed with all Radiomics 1−13 and their coefficients. Next, the lung radiomics feature Y2 is constructed with Radiomics 1−3 and their coefficients belonging to the original lung parenchyma images. Then, the lung radiomics feature Y3 is constructed with Radiomics 4−8 and their coefficients belonging to the derived lung parenchyma images generated from the Log filter. Finally, the lung radiomics feature Y4 is constructed with Radiomics 9−13 and their coefficients belonging to the derived lung parenchyma images generated from the wavelet filter. Figure 8 shows the significance of the lung radiomics features Y1−Y4 among different COPD stages. Table 4 shows the adjusted P-value of the lung radiomics features Y1−Y4 among different COPD stages.

Figure 8.  The changing trend of the lung radiomics features Y1−Y4 among different COPD stages using Dunn's multiple comparisons test. Figure 8(A) shows the changing trend of the lung radiomics feature Y1 constructed by all the 13 selected COPD radiomics features. Figure 8(B) shows the changing trend of the lung radiomics feature Y2 constructed by Radiomics 1−3. Figure 8(C) shows the changing trend of the lung radiomics feature Y3 constructed by Radiomics 4−8. Finally, figure 8(D) shows the changing trend of the lung radiomics feature Y4 constructed by Radiomics 9−13.

DownLoad: Full-Size Img PowerPoint

Table 4.  The adjusted P-value of the lung radiomics features Y1−Y4 among different COPD stages using Dunn's multiple comparisons test.

COPD stage comparison	Y1	Y2	Y3	Y4
Stage 0 vs. Stage Ⅰ	0.0542 (ns)	0.0641 (ns)	> 0.9999 (ns)	< 0.0013
Stage 0 vs. Stage Ⅱ	< 0.0001	0.1686 (ns)	0.0686 (ns)	< 0.0001
Stage 0 vs. Stage Ⅲ & Ⅳ	< 0.0001	> 0.9999 (ns)	> 0.9999 (ns)	< 0.0001
Stage Ⅰ vs. Stage Ⅱ	0.0718 (ns)	> 0.9999 (ns)	0.3835 (ns)	0.1287 (ns)
Stage Ⅰ vs. Stage Ⅲ & Ⅳ	< 0.0001	0.0729 (ns)	> 0.9999 (ns)	< 0.0001
Stage Ⅱ vs. Stage Ⅲ & Ⅳ	< 0.0039	0.1851(ns)	0.1143 (ns)	0.0619 (ns)
*Note: ns: no significance.

---

 | Show Table

DownLoad: CSV

Table 4 shows no significant change in lung radiomics feature Y2 and Y3 among the COPD stages. Therefore, the lung radiomics feature Y1 and Y4 are further analyzed. Figure 8(A) shows that lung radiomics feature Y1 increases with the development of COPD stages. The mean ± SEM of lung radiomics feature Y1 from the COPD stage 0 to Ⅲ & IV is −0.1960 ± 0.0303, −0.0507 ± 0.0268, 0.04254 ± 0.0201 and 0.2302 ± 0.0311, respectively. However, there is no significant change between COPD stage 0 and COPD stage I (P = 0.0542) and between COPD stage I and COPD stage Ⅱ (P = 0.0718) shown in Table 4. Like lung radiomics feature Y1, Figure 8(D) shows that the lung radiomics feature Y4 also increases with the development of COPD stages. The mean ± SEM of lung radiomics feature Y4 from the COPD stage 0 to Ⅲ & IV is −0.2234 ± 0.0320, −0.0272 ± 0.0235, 0.0651 ± 0.0220 and 0.2058 ± 0.0327, respectively. There is no significant change between COPD stage I and COPD stage Ⅱ (P = 0.1287) and between COPD stage Ⅱ and COPD stage Ⅲ & IV (P = 0.0619) shown in Table 4.

Two lung radiomics features, F1 and F2, are constructed using the other COPD radiomics combination strategy seen in Table 5. Figure 9 shows the changing trend of the lung radiomics features F1 and F2 constructed by the COPD radiomics features combination strategy (Section 2.2.4) among different COPD stages. Figure 9(A) shows the changing trend of the lung radiomics feature F1 constructed by Radiomics 13 with the coefficient −0.2494 and Radiomics 3 with the coefficient −0.0715. Figure 9(B) shows the changing trend of the lung radiomics feature F1 constructed by Radiomics 13, Radiomics 3, and Radiomics 12 with the coefficient −0.2494, −0.0715, −0.0590, respectively. Figure 9(A) and Table 5 show that except for the significance between COPD stage Ⅰ and Ⅱ (P = 0.0646), the lung radiomics feature F1 significantly increases with the development of COPD stages. The mean ± SEM of the lung radiomics feature F1 from the COPD stage 0 to Ⅲ & IV is −0.1990 ± 0.0259, −0.0433 ± 0.0218, 0.0398 ± 0.0177, and 0.2294 ± 0.0268, respectively. Figure 9(A) and Table 5 also show that the lung radiomics feature F2 significantly increases with the development of COPD stages. The mean ± SEM of the lung radiomics feature F2 from the COPD stage 0 to Ⅲ & IV is −0.1739 ± 0.0255, −0.0502 ± 0.0215, 0.0521 ± 0.0183, and 0.1901 ± 0.0280, respectively.

Figure 9.  The changing trend of the lung radiomics features F1 and F2 among different COPD stages. Figure 9(A) shows the changing trend of the lung radiomics feature F1 constructed by Radiomics13 and Radiomics3. Figure 9(B) shows the changing trend of the lung radiomics feature F1 constructed by Radiomics 13, Radiomics 3, and Radiomics 12.

DownLoad: Full-Size Img PowerPoint

Table 5.  The adjusted P-value of the lung radiomics features F1 and F2 among different COPD stages using Dunn's multiple comparisons test.

COPD stage comparison	F1	F2
Stage 0 vs. Stage Ⅰ	0.0031	0.0479
Stage 0 vs. Stage Ⅱ	< 0.0001	< 0.0001
Stage 0 vs. Stage Ⅲ & Ⅳ	< 0.0001	< 0.0001
Stage Ⅰ vs. Stage Ⅱ	0.0646 (ns)	0.0085
Stage Ⅰ vs. Stage Ⅲ & Ⅳ	< 0.0001	< 0.0001
Stage Ⅱ vs. Stage Ⅲ & Ⅳ	< 0.0014	0.0342
*Note: ns: no significance.

---

 | Show Table

DownLoad: CSV

4.   Discussion

The selected COPD radiomics features, the lung radiomics features Y1−Y4, and the lung radiomics features F1−F2 related to the resting HR are discussed in Figures 6−9 and Tables 3−5.

From the above results, a single selected COPD radiomics features related to the resting HR cannot characterize the significant changes of COPD stage evolution among different COPD stages, especially from COPD stage I to COPD stage Ⅱ. Although Radiomics 1 and Radiomics 13 can reflect significant changes in most COPD stages, they only fail to characterize the significant changes from COPD stage I to COPD stage Ⅱ. Compared to Radiomics 1, the significance from COPD stage I to COPD stage Ⅱ in Radiomics 13 (P = 0.4088) is better than that of Radiomics 1(P > 0.9999). Radiomics 3 with significance (P < 0.0001) is the sensitive COPD radiomics feature related to resting HR from the risk of COPD (COPD stage 0) to suffering from COPD (COPD stage I−Ⅲ & IV). At the same time, the significance of Radiomics 13 among other COPD stages is also better than that of Radiomics 1. Most importantly, Radiomics 13 is the dominant COPD radiomics feature affecting the resting HR. Therefore, the dominant COPD radiomics feature Radiomics 13 not only characterizes the resting HR but also characterizes COPD stages evolution (except for the significance between COPD stage I and COPD stage Ⅱ).

The selected COPD radiomics features obtained by different types of lung parenchyma images are also further discussed in this paper. Unfortunately, the lung radiomics feature Y2, calculated by the original lung parenchyma images, fails to characterize the evolution of COPD stages. Likewise, the lung radiomics feature Y3, calculated by the derived lung parenchyma images (LoG filters), only characterizes the COPD stage 0 to COPD stage Ⅱ. However, the lung radiomics feature Y4, calculated by the derived lung parenchyma images (wavelet filter), only fails to characterize the COPD stage I to COPD stage Ⅱ. Therefore, the selected COPD radiomics features calculated based on wavelet filter can better characterize COPD stage evolution than the LoG filter. At the same time, it can be seen from the lung radiomics feature Y1, which fully characterizes the resting HR, that it improves the overall significance among the different COPD stages. In particular, Radiomics 13 is also calculated from the derived lung parenchyma images based on the wavelet filter. The lung radiomics feature F1 and F2 are also discussed. The lung radiomics feature F2 constructed by Radiomics13, Radiomics 3, and Radiomics12 improves the significance from COPD stage I to COPD stage Ⅱ (P = 0.0085 < 0.05). Although Radiomics13 is the dominant COPD radiomics feature, other selected COPD radiomics features Radiomics 3 and Radiomics 12 are also needed to characterize the COPD evolution. No matter what, finding a lung radiomics feature is a competitive process between different COPD stages.

The COPD radiomics features are calculated based on chest HRCT images with different COPD stages. The selected COPD radiomics features related to the resting HR are further determined. The selected COPD radiomics features can reflect the resting HR variability. Therefore, the relationship between the COPD radiomics features and the resting HR is revealed. The selected COPD radiomics features and/or the lung radiomics feature F2 may be a predictor of the resting HR variability of the subjects with COPD. The resting HR at different COPD stages may be predicted by the selected COPD radiomics features and/or the lung radiomics feature F2. Clinically, the resting HR variability has many causes. Although the abnormal resting HR caused by other diseases and outside the interval ^[5] was excluded in our study, it will also be more meaningful to analyze the patients with cardiopulmonary disease in future research.

The COPD radiomics features are calculated from the lung parenchyma images, reflecting the state of "COPD patients" at stages 0−IV. Compared with chest HRCT images, the COPD radiomics features can express the hidden information at different COPD stages. This hidden information is more helpful to characterize the differences of varying COPD stages and releases the relationship between resting HR and COPD evolution. There are also some limitations of this study. First, when chest HRCT images are collected, the patient's inspiratory state can be controlled, but the heart's movement cannot be controlled. Therefore, the blood state in pulmonary vessels must impact the calculation of the COPD radiomics features. Second, the number of subjects with COPD stage IV is only 10. Therefore, we take COPD stage Ⅲ and IV as one COPD stage, affecting the analysis of the results.

5.   Conclusions

Massive COPD radiomics features are calculated based on the lung region segmented by the trained ResU-Net. The 13 selected COPD radiomics features related to the resting HR are selected from the massive COPD radiomics features using the Lasso model. A COPD radiomics features combination strategy is proposed to provide a lung radiomics feature for characterizing the resting HR and the COPD stage evolution. Because the lung radiomics feature Y1 is constructed by all the selected COPD radiomics features, it is considered that it can fully characterize the resting HR. However, the lung radiomics feature Y1 fails to characterize the COPD stage evolution from COPD stage 0 to I, and from COPD stage I to Ⅱ. Compared with the P-value of Y1, that of the lung radiomics feature F2 has been improved 0.63% between COPD stage 0 and stage Ⅰ, and 6.33% between COPD stage Ⅰ and stage Ⅱ, resulting in the P-value of F2 less than 0.05. Based on the COPD radiomics features combination strategy, the lung radiomics feature F2 with the dominant selected COPD radiomics features not only can characterize the resting HR but also can characterize the COPD stage evolution.

Acknowledgments

Thanks to the Department of Radiology, the First Affiliated Hospital of Guangzhou Medical University for providing the dataset and the funding support from the National Natural Science Foundation of China (62071311), Natural Science Foundation of Guangdong Province, China (2019A1515011382), Stable Support Plan for Colleges and Universities in Shenzhen, China (SZWD2021010), and Scientific Research Fund of Liaoning Province, China (JL201919).

Conflict of Interest

The authors declare no conflict of interest.