Potential role of CT-textural features for differentiation between viral interstitial pneumonias, pneumocystis jirovecii pneumonia and diffuse alveolar hemorrhage in early stages of disease: a proof of principle

This was a retrospective CT, clinical (microbiological) and BAL-data evaluation which was approved by the local ethic committee. By retrospective database search of the local radiology department and bronchoscopy centre we identified 62 suitable patients. Because of missing CT examinations we must exclude 16 patients, so that finally 46 patients were included (Fig. 1). These 46 consecutive patients (female, 17; male, 29; mean age 62.70y ± 14.02 y; range, 29–85 y) with RSV (n = 5), HSV1 (n = 6), PJP (n = 21) and lung hemorrhage (n = 14) underwent high-resolution chest CT within 10 days after institution of respiratory symptoms between January 2016 and February 2017. Underlying diseases and patient characteristics were shown on Table 1.

Mean time between bronchial lavage and CT imaging examination was 11.18 ± 1.61 days.

Patients were retrospectively recruited for both HRCT- and CTTA-analysis if they fulfilled the following inclusion criteria: 1) positive bronchoscopy for pulmonary haemorrhages, viral pneumonia or Pneumocystis jirovecii pneumonia; 2) at least one HRCT of the lung at the onset of the disease; 3) age over 18 years. 16 patients had to be excluded as they had no CT-diagnosis. Exclusion criteria were additional pathologies affecting the lung parenchyma and overlying the proposed clinical pathologies: 1) pleural effusions; 2) lung edema or; 3) additional bacterial infections. The process of patient recruitment is shown on Fig. 1.

According to BAL-analysis, 5/46 (10.8%) patients had RSV, 6/ 46 (13.0%) had HSV1 and 21/46 (45.6%) had PJP. Alveolar hemorrhage was diagnosed by BAL in 14/46 patients (30.6%).

Assignment of the patients to one of the three categories viral or PJP pneumonia and diffuse alveolar hemorrhage was based on microbiological data collected by BAL or by evidence of blood in the bronchial lavage. Diagnosis of herpes simplex virus pneumonia was based on the isolation of the virus by cell culture. Monolayers of human foreskin fibroblasts and vero cells were inoculated with bronchoalveolar lavage (BAL) and maintained in culture for up to 2 weeks. The virus was identified by its characteristic cytopathic effect and immunoperoxidase staining for HSV glycoprotein D.

Detection of respiratory syncytial virus (RSV) from BAL was done by real time PCR using a commercially available assay according to the instructions of the manufacturer (RealStar RSV RT-PCR Kit, altona Diagnostics GmbH, Hamburg, Germany). All patients with fresh alveolar hemorrhage had blood in the BAL.

All scans were viewed at standard lung window (level, − 700 HU; width, 1500 HU by two independent readers (MH, CK) with 25 and 5 year experience in reading chest-CT. Image data was first analysed for the presence of lung parenchymal abnormalities (ground-glass opacity-GGO, crazy paving, air-space consolidation, reticulation, thickening of the bronchial wall and centrilobular nodules-(tree-in-bud)) including their anatomical distribution (central vs. peripheral) and predominance. Ground-glass attenuation was defined as an area of hazy increased attenuation without obscuration of underlying vascular markings. Air-space consolidation was considered present when the opacities obscured the underlying vessels. Crazy-paving was considered when parenchymal attenuation reached intermediate values between GGO and air-space consolidation enabling distinction of vascular and bronchial structures. Reticulation was defined as thickening of interlobular and intralobular septa. The presence of centrilobular nodules (tree-in-bud) as well as thickening of the bronchial walls (bronchial cuffing) was also registered. Furthermore, anatomic distribution was regarded as predominantly peripheral if abnormalities were seen mostly in the outer third of the lung vs. predominantly central if most were in the inner third of the lung. Finally, the incidence of all HRCT-imaging findings was quantified determining the predominant involvement pattern. The extent of pulmonary disease was evaluated in 3 levels of the lungs, as follows: the upper level was defined as the area above the level of the carina; the middle level, between the level of carina and the level of the inferior pulmonary veins; and the lower level, beneath the level of the inferior pulmonary veins. Each HRCT sign was separately coded as present or absent in these six areas. For each zone, a score was based on a visual estimation of the percentage of lung tissue demonstrating the CT sign (1, extent < 25%, 2: extent between 25 and 50%; 3: extent between 50 and 75%; and 4: extent > 75%). The global extent score of each sign was the sum of the 3 zonal scores [12].

HU measurements of parenchymal abnormalities were not performed as they are indirectly reflected by the mean intensity and mean average values obtained at CTTA.

In the next step large volumes of interest (VOI) were set in the most representative areas in both lungs (if both involved) in order to perform texture analysis (Fig. 2).

CTTA was evaluated using a prototype software (Siemens Healthcare) based on image data sets of 1 mm slice thickness. HU units between − 600 and + 200 HU were involved [13]. The computation of each texture type for an input volume of interest (VOI) involved assigning a new value (“texture value”) to all voxel of that VOI. This technique includes an image filtration step to selectively extract features of different sizes and intensity variation, followed by texture quantification. Series of derived images at different spatial scales from fine to coarse texture within a volume of interest (VOI) drawn in lung.

The VOIs were positioned on typical disease manifestations both in the right and the left lung. A total of 82 VOIs were drawn, 41 in the left lung and 41 in the right lung. VOIs were placed only in the left or right lung in case of unilateral disease involvement. The areas were the VOIs were set was decided by the senior reader (25 years’ experience in thoracic imaging) and represented the predominantly involved lung areas and CT-features.

As we addressed a mixed of viral and fungal pneumonias as well as lung hemorrhage in early stages, filter parameter were tuned for medium texture. A computation was performed on the current voxel and its neighbourhood, and the results of that were stored as the current voxel’s texture value. This was repeated for every voxel in the VOI. All computed CTTA-parameters and their meaning are shown on Table 2.

All results are expressed as average with standard deviation. Statistical analysis was performed using dedicated software (IBM SPSS 22.0 (SPSS, Armonk, USA). The Kolmogorow-Smirnov test was used for the normality test including Lilliefors significance correction. For correlation Pearson’s correlation coefficient were calculated.

Incidence of GGO, crazy-paving, air-space consolidation, centrilobular nodules/tree-in-bud, thickening of bronchial wall and reticulation is presented on Table 3. No statistical significant differences were registered for GGO (p = 0.703), crazy-paving (p = 0.0509), centrilobular nodules (p = 0.997) and parenchymal consolidations (p = 0.478) between the different subgroups.

Statistical significant differences were registered for reticular consolidations (significant more often in alveolar haemorrhages, p = 0.012) and distribution pattern of peripheral and central lung pattern (p = 0.021). According to this data, peripheral distribution was significantly more often registered in viral pneumonia than in alveolar haemorrhages and PJP. Exact distribution is shown on Table 3.

A subpleural sparing was found in 12/21 PJP, 6/14 alveolar hemorrhages and 4/11 viral pneumonias.

Discriminatory CTTA-features were found between diffuse alveolar hemorrhage and PJP consisting of differences in mean heterogeneity (p < 0.015) and uniformity of skewness (p < 0.006). Statistically significances were tabulated on Table 4. Detailed value results of CTTA-analysis were shown on Additional file 1: Table S1.

There was no difference between CT-textural features of diffuse alveolar hemorrhage and viral pneumonia or PJP and viral pneumonia.

a) Alveolar hemorrhages (A).

Crazy paving correlated significantly with mean intensity (p = 0.016, r = 0.630), entropy of intensity (p = 0.013, r = 0.644), mean average (p = 0.013, r = 0.642), entropy of average (p = 0.004, r = 0.711) and uniformity of average (p = 0.025, r = − 0.595).

b) PJP (B).

Centrilobular nodules correlated with entropy of intensity (p = 0.013, r = − 0.607) and uniformity of intensity (p = 0.019, r = 0.577). Furthermore the presence of centrilobular nodules correlated with entropy of average (p = 0.016, r = − 0.589) and uniformity of average (p = 0.018, r = 0.583).

c) Virus pneumonia (C).

Results of fine and course filter were added as Additional file 1: Table S1.

Chest-CT manifestations also significantly overlap with all six imaging features being present in all our patients. Notably, the mean intensity in alveolar hemorrhage was not increased as expected. This is presumably due to the early phase of hemorrhage where the hematocrit is still low and thus the attenuation remains in the range also expected from air-space consolidation. According to visual CT-assessment, in PJP GGO generally predominated whereas in alveolar hemorrhage air-space consolidation was more frequent which in our opinion explains differences in tissue heterogeneity and the entropy of skewness between these two pulmonary complications.

The potential benefit of using CT-textural features for characterization of lung parenchymal abnormalities has been already demonstrated in previous reports [10, 11, 18]. The main strength of this technique consists in its histogram-based ultrastructural tissue analysis which enables to some extent parallels macro-histology. This is particularly true in the lung which has a predictable three-dimensional texture. Behind 1st order statistical features delivering information about tissue intensity, structural heterogeneity and or uniformity, 2nd order statistics give a more profound insight into the tissue buildup analyzing the matrix (each single voxel) and deriving sensitive data with respect to their spatial distribution. In the past, numerous texture features characterizing grey-level intensity and distribution have been correlated with the tumor microarchitecture defined by histology [13, 19]. Some few reports have already addressed the issue of using CT textural analysis for characterization of lung expanding its application area [11, 20]. Park et al. found a better correlation of texture-based quantification of lung emphysema with PFT compared to density-based evaluation [10]. Cunliffe et al. applied CT textural analysis for monitoring the course of radiation-induced pneumonitis defining specific CT textural analysis features capable to accurately accomplish this task [11, 21]. However, the prerequisite for a successful performance of this technique is a histology-based differentiation between normal and abnormal lung parenchyma on the one side and also between pulmonary abnormalities to be characterized on the other side. At this point, our data confirms the potential of CT-textural features for tissue characterization, but also its limits in case of similar pathologic tissue changes as with the disorders that we have addressed in our study.

The different imaging features detected by visual evaluation of HRCT-scans showed some good correlations with CT-textural features. Hence, crazy-paving in cases with alveolar hemorrhage correlated positively with mean intensity and entropy of intensity as well as with mean average and entropy and uniformity of average (noise independent voxel intensity) and negatively with the uniformity of average. This is presumably a measure for the distribution of parenchymal attenuation in this pulmonary complication.

In viral pneumonias, reticulation correlated well with the uniformity of entropy of co-occurrence matrix which represents the distribution of neighbors (voxel with a certain grey value) in a CT-morphological lattice-like attenuation pattern. Finally, the presence of centrilobular nodules in PJP correlated positively with the uniformity of intensity and average and negatively with the entropy of intensity and average. This shows how a punctiform attenuation pattern affects the relationship between neighbors in terms of regularity of intensity and probability of distribution.

This study has some limitations. First, it was retrospective in character. Second, not all our patients were in the clinical setting of immunosuppression before or shortly after allogeneic bone marrow transplantation which was the hypothetical scenario for the rationale of such a project. Nonetheless, the amount of e.g. infection-related pulmonary changes is not only dependent on the immune status of affected patients but also by the virulence of the pathogen. Similarly, imaging features of alveolar hemorrhage occurring in the setting of overdosed anticoagulant drugs or due to leukemia or temporarily post-treatment thrombocytopenia are expected to exhibit similar histologic features. Third, the number of cases retrospectively included was small for a robust statistical evaluation.

Nonetheless, we consider that research on this topic should be extended and focussed on CTTA-differences between HRCT-image findings irrespective of their cause aiming at automated detection and classification of such features by using machine learning algorithms. Our report is only a proof of principle, further extended studies is necessary.