Background
“Radiomics” is an approach currently recognized in biomedical image analysis as a tool to define a potentially diverse array of meta-data obtained from images using quantitative radiology image analytics. Some well-selected features of these meta-data can be informative of the health status of the imaged organ system and impact therapy decisions. Such therapy decisions are best taken early in the course of disease. Early therapy decisions have tremendous impact on quality of life in pulmonary diseases.
Fractals are often used to characterize non-Euclidean structures in biology [
1]. Utilizing the scaling factor of statistically self-similar and non-overlapping subsets, fractal dimension can be computed [
2] providing relevant information describing a structure’s complexity and homogeneity [
3]. Fractal dimension represents, with certain limitations, the “less or more branching nature” of structures [
4,
5] including the respiratory organ [
6].
We sought to implement a fractal-based radiomics approach to X-ray computed tomography attenuation data without respiratory gating thereby averting the risk of losing relevant information. Analysis of fractal dimensions in specially binned non-gated X-ray computed tomography image patterns has been the method of our choice.
The currently accepted method of analysing pulmonary fractal dimensions of X-ray computed tomography attenuation data usually consists of segmenting parts of the lung such as the alveolar respiratory units, or pulmonary arteries and veins [
7‐
10]. Al-Kadi and Watson [
2] distinguished tumours and blood vessels based on their X-ray attenuation differences (using contrast material) to perform fractal dimension analysis on the image segments. The usually applied methods therefore provide the reader with a fractal dimension value for each of the tissue component segments of lung images. A usual outcome measure e.g. is the fractal dimension of lung arterial vasculature.
Our approach to radiomics has been greatly different from simply calculating fractal dimensions of segmented pulmonary tissue components (“dissected” vessels, bronchi etc.). The examination of fractal dimensions of ideally selected attenuation ranges in relative Hounsfield units (HU) may provide the foundation towards discovering additional hidden tissue features in integrative patterns of lung images instead. These fractal dimension calculations may perhaps detect small scale tissue alterations such as those caused by harmful environmental conditions. Our objective was to unveil possible correlations between air pollutant categories and specific features or patterns of damaged lungs. These features of small magnitude might not be evident in either custom visual X-ray computed tomography image analysis or in the calculation of segmented pulmonary tissue fractal dimensions.
We aimed at distinguishing between different air pollutant effects on the lungs via a radiomic approach with a clinically translatable mathematical algorithm. We preferred using non-gated X-ray computed tomography data. Non-gated data acquisition still contains effects of e.g. hindered chest or lung motion. In our analysis we intended to examine data features reflecting disease-related changes also in lung organ movements rather than anatomical relationships. Thus in our analysis method presented here simple non-gated X-ray computed tomography mouse chest scans have been acquired and evaluated by the calculation of fractal dimension of binary images. We binned voxel sets from each mouse chest X-ray computed tomography volume into numerous attenuation ranges in our study [
1,
11], instead of pulmonary tissue-based image segmentation. Additionally we also averted the use of any contrast agent.
Generally speaking, (both in the “classic” and in our novel method), the result of fractal dimension calculation is a number corresponding to how often examined structures (dissected arteries and veins in “classical” methods and voxel 3D patterns with specific attenuation values in our approach) branch and/or fill the space within the chest. However, in our novel approach, the fractal dimension calculation method examines and depicts integrative binary images of lung voxels which are selected according to their attenuation values. We then aimed at the application of our algorithm to discriminate among groups of mice treated with different air pollutants in an early phase of their respective disease models.
Results
The mean values of width and position of the voxel density histograms demonstrate no significant differences between the three groups (Fig.
2).
The fractal dimension - cut-off range functions were evaluated by fitting them with Gaussian curves “A” and “B” (Fig.
4). These functions can be characterized by height, maximum position and width of the peak. The means of height, width and position of the “A” curve of the CON group do not differ significantly when compared to the SDO or the SAO group. The means of height, width and position are unchanged between the SDO and SAO groups, too (Fig.
5a).
The mean of height of “B” curve of the SDO group increased significantly when compared to the CON group (KW
p = 0.002, MWph
p = 0.036) and to the SAO group (KW
p = 0.002, MWph
p = 0.024), but not significantly when the CON group was compared to the SAO group (Fig.
5b top). The mean of widths of “B” curves of the SDO group increased significantly (KW
p = 0.016, MWph
p = 0.036) compared to the CON group, and also significantly when compared to the SAO group (KW
p = 0.016, MWph
p = 0.024), but not significantly when the CON group was compared to the SAO group (KW
p = 0.016, MWph
p = 0.429) (Fig.
5b middle). The means of maximum positions are not significantly altered between the SDO, SAO and CON groups (Fig.
5b bottom).
The difference between the ratios of height is significant if the SDO group is compared to the CON group (KW
p = 0.005, MWph
p = 0.0357), if the SDO group is compared to the SAO group (KW
p = 0.005, MWph
p = 0.024) and if the SAO group is compared to the CON group (KW
p = 0.005, MWph
p = 0.042) (Fig.
5c top).
The ratios of width of the SDO and CON groups (KW
p = 0.021, MWph
p = 0.036) and the SDO and SAO groups (KW
p = 0.021, MWph
p = 0.024) demonstrate a slight but significant difference (Fig.
5c middle), however, the difference between the SAO and CON groups is not significant.
Discussion
Differences between parameters based on the relative HU--frequency histograms of the animals in the three groups, i.e. the differences between mean maximum positions and widths (Fig.
2) could be neither the basis of detection of an altered lung structure nor the categorization of air pollution exposure. Either the absence of gating or the reduced period of mouse model symptom production was likely the reason. Indeed, Sasaki et al. (2015) too could not distinguish the effect of cigarette smoke on lung tissue attenuation in comparison to the control using X-ray computed tomography with gating [
15]. However, interestingly enough, a number of differences could be shown using our novel radiomics analysis method.
In our opinion, there are distinctly altered tissue features with respectively changed attenuation patterns in the different mouse models of air pollution related disease. However, the readout of these changes necessitates subtle differentiation and radiomics analysis methods in early phases of lung harm when symptoms are yet considerably subtle.
Severe lung diseases alter the inhalatory and exhalatory movements and these changes can be detected even in mild cases or after relatively short exposure to air pollutants (e.g., the slower exhalation in COPD is evident) [
16,
17]. This may be paradoxically advantageous in our approach and likely will help in discerning exposure sources. The hindered motion in the smoke/air and sulphur dioxide groups presumably decreases or even negates spatial and temporal overlapping of different tissues in the same voxel caused by respiratory movement and finally contributes to the mentioned increase of width parameter in the exposed groups. This change in motion dynamics (probably due to inflammation, mucus build-up and entrapped air bubbles) may be an important part of the diagnosis and it is ignored when using respiratory gating. Notably, this finding suggests the fractal dimension- cut-off range function derived radiomic data might unveil some pathologic changes in lung diseases [
3,
18]. Possibly the onset of disease as a result of exposure to air pollution could also be observed with our data analysis method.
The Gaussian curves of Fig.
5 display a different pattern of respiratory pulmonary motion in the exposed animals, possibly due to an increase in lung stiffness caused by pollutants.
The height of Gauss curve B is significantly increased in the SDO group compared to the other two groups (Fig.
5b top). We infer this change is attributable to the hindered motion of inflamed tissue.
Sulphurous gases are irritants and induce inflammation, bronchoconstriction and bronchitis resulting in an increase of mucus [
3,
19]. Overproduction of mucus can form plugs which entrap air or temporally and partly obstruct the upper airways [
20]. Airway clearance of mucus depends on the interactions between physical properties of the mucous gel, serous fluid content, and ciliary function, in addition to airflow [
19]. Wagner et al. (2006) discovered in a Sprague–Dawley rat model that 80 ppm concentration of SO
2 (besides overproduction of mucus) caused epithelial cells to lose their ciliae [
21]. Nano-sized solid particles originating from fumes tend to accumulate in deeper airways and alveoli [
22], as inflammatory agents increase water permeability and dilate cell volume thus thickening airway walls and resulting in the narrowing of the airways (in addition to minor mucus production which cannot be excluded). In our interpretation, this narrowing of the airways causes the different motion dynamics of this group.
The width parameter of the SDO group is significantly increased compared to the other two groups (Fig.
5b center). We believe this change is attributable to the hindered motion of inflamed tissue.
The number of voxels representing thickening airway walls is increased, caused by SO
2 exposure often penetrating into deeper airways and inducing inflammation in the alveoli, leading to the appearance of fluid, derived from necrotic cells [
19,
23]. Mucus plugs trap air inside the alveoli and lead to the formation of micro-sized bubbles [
24,
25] inside the lung parenchyma. Indeed mucus production is an early response to increased amounts of air pollution [
19]. In our interpretation, the increased number of voxels representing thickening airway walls causes the different shape of the fractal dimension - cut-off range function of this group.
Only a slight difference was observed between the means of height of Gaussian curve “B” of SAO and CON groups (Fig.
5b top). In addition to mean values of maximum positions (Fig.
5b bottom), the width parameter of the SAO group was compared to the control group (Fig.
5 middle), however it did not significantly change.
Theoretically speaking, emphysema-diseased areas within the lungs could occur caused by destroyed walls of airways and alveoli [
11,
12,
18], however, it appears only in long term experiments [
26].
In using the corresponding ratios of parameters of Gaussian curves “A” and “B” (Fig.
5c) and the heights and widths of “B” Gaussian curves (Fig.
5b), all three groups could be distinguished. We believe the difference between the ratios of parameters refers to the different proportion of various kinds of tissue damage caused by different air pollution agents.
The proportion of tissue alterations has a specific pattern in lung diseases of different origin. Altered tissue (e.g., increased mucus production in one model or increased presence of tissue microbubbles in another) was explored.
The mechanism of these changes affecting the respiratory movement remains unclear and warrants further research.
Data acquired in our study (Fig.
5b, c) proved a worthy basis for differentiating specific air pollution caused lung changes in the early stage by direct fractal dimension - cut-off range function pattern analysis. It could be hypothesized that molecular features and presence of mucus in smaller airways [
19] and inflammation profile of lung tissues [
27] contribute to our fractal dimension analysis-based results. In reference to published literature [
19,
28], we postulate that the fractal dimension - cut-off range function calculated with our method may be used as an imaging biomarker. It could be effectively converted to both preclinical applications and clinical use in humans, ideally providing patients the benefit of early warning towards avoiding environmental risks. Additional benefits are expected in the proper treatment at the onset of symptoms of disease, and lastly, to prevent aggravation of disease and exacerbation of COPD and/or asthma bronchiale symptoms. Here, we highlight the importance of length of smoke exposure and genetic susceptibility [
28] to emphysema [
26], which may later develop and will be reflected in the reported imaging biomarker parameters [
3,
18].
The radiomic analysis of the fractal dimension - cut-off range function may be useful in early diagnosis of both exposure to air pollution and lung diseases (such as COPD or asthma), containing information about both the molecular features and patterns of mucus in smaller airways [
19] and the inflammation profiles of lung tissues [
27]. We propose an increase in the number of structurally quantitative imaging biomarker research studies, towards early detection and follow-up of therapy of other pulmonary diseases, for example, cystic fibrosis, lung carcinomas [
3] or tuberculosis [
29].
In translational research, our most important goal is to develop methods in animal models which later may be used in clinical practice. In the case of our paper, the translatability of the method is dependent upon three aspects of it. The first is the usability of the algorithm in clinical practice, the second is the relation between anatomy sizes and reconstruction voxels, and the third is the applicability of the algorithm in clinical protocols.
The algorithm does not use specific data, as it only requires a 3D reconstruction, that is attenuation distribution. These data are also available in the clinical setting, so our algorithm proposed here can also be applied for clinical lung computed tomography volumes.
The human body is about 15–20 times longer than the body of the mouse and there is nearly the same difference between spatial resolution (voxel size in preclinical CT is 50 μm and in clinical CT is 500–750 μm). Generally speaking, anatomy size and voxel size change proportionally. Because of the partial-volume effect, though, true resolution does not reach voxel size calculated from reconstruction. It is important to note that the size of alveoli is invariable between species and is around 200 μm. Consequently, neither the pre-clinical nor the clinical instrument can visualize individual alveoli with suitable resolution, but in the case of the pre-clinical instrument, we see one alveolus in the adjacent voxels, whereas in case of the clinical instrument, we see more than one alveoli in one voxel.
Practically speaking, the proposed algorithm can be applied in the clinics and it is not necessary to change protocols, since the usual examinations and unchanged data acquisition chain should be followed by this new “off-line” analysis.
In summary, the use of the algorithm will be self-evident in clinical practice. Naturally, its diagnostic effectiveness needs to be assessed meticulously throughout different diseases.
Competing interests
The authors hereby declare they have no competing interests. Tibor Szabó, Ilona Czibak, Márta Pócsik, and Domokos Máthé are employees of CROmed Translational Research Centers. CROmed Translational Research Centers does not have any financial or non-financial interest in the subject matter or materials discussed in this manuscript. Ferenc Budán is an employee/stakeholder of MedProDevelop Kft that has no financial or non-financial interest in the subject matter or materials discussed in this manuscript.
Authors’ contributions
DM and TS designed the study; TS, IC, DM and IH did the preclinical imaging work; CK, DSV and FB analyzed the data; KS takes responsibility for the integrity of the data and the accuracy of the data analysis; DM, KS, FB, ZG and MP wrote the paper; KK and RB contributed expert advice and valuable insight. All authors read and approved the final manuscript.