Assessing the impact of CT reconstruction kernel on radiomic features extracted from normal and fibrotic tissue in patients with diffuse lung disease

Abstract

Introduction:

Patients with diffuse lung disease (DLD) undergo CT scans for diagnosis and evaluation. Attempts to characterize the radiographic appearance of these regions with quantitative features like radiomics have been hindered by variability in CT acquisition and reconstruction parameters. The purpose of this investigation is to characterize the effect of CT reconstruction kernel on radiomic features in normal and fibrotic regions in DLD patients.

Methods:

Raw CT projection data of DLD patients receiving a thoracic CT exam was collected from 3 CT scanners (Definition AS, Flash, and Force; Siemens Healthineers, Forchheim, Germany) and retrospectively reconstructed with 5 reconstruction kernels (smooth, medium-smooth, medium, medium-sharp, and sharp). The medium kernel is part of the clinical protocol at our institution and considered the reference kernel for this investigation. Regions of classic normal and classic fibrosis were annotated by 2 thoracic radiologists on images reconstructed with reference kernels. The annotations were copied across each reconstruction, and 72 radiomic features were extracted from each annotation using Pyradiomics. Agreement between features in reference and non-reference kernels was assessed with concordance correlation coefficient (CCC). Features were considered robust if average CCC was > 0.9.

Results:

There were 66 patients included with 116 regions of normal and 208 regions of fibrosis annotated across all patients. Across kernels in normal tissue, 0/16 GLSZM, 1/16 GLRLM, 0/22 GLCM, and 5/18 first-order features were robust (average CCC > 0.9). Across kernels in fibrotic tissue, 1/16 GLSZM, 3/16 GLRLM, 3/22 GLCM, and 7/18 first-order features were robust, however the effect magnitude of sharper kernels was greater than in normal tissue.

Conclusion:

First-order features were more robust than other features; GLCM features were the least robust. There are more robust features across kernels in fibrotic tissue, but the magnitude of kernel effect is greater in fibrotic tissue than in normal tissue.

1. INTRODUCTION

Diffuse lung disease (DLD) is a group of pulmonary diseases involving inflammation of lung parenchyma that mechanically supports airways, vascular structures, and alveoli.¹ DLD manifests in a variety of pathologies, symptoms, and prognoses, however all forms are capable of causing irreversible pulmonary fibrosis (PF) and are associated with increased risk of pulmonary hypertension and lung cancer.^2–4 More than 30% of patients develop PF and have a median survival of less than 5 years after diagnosis.^5,6 Currently, there are FDA approved treatments that slow progression, however, there is no known cure.

Patients with diffuse Lung Disease (ILD) frequently undergo CT scans for diagnosis and evaluation. Radiologists typically evaluate the extent of different types of radiographic appearance such as regions of fibrosis, honeycombing, and ground glass attenuation; these regions are often characterized by their attenuation and texture appearance. Quantitative image analysis can augment subjective, expert interpretation of medical image data with objective, quantitative information using radiomics. However, attempts to characterize these regions with quantitative imaging features such as radiomics have been hindered by variability in CT acquisition and reconstruction parameters.^7,8 Numerous studies in phantoms^9–13 and patients^8,14–17 have demonstrated the impact of CT scanner and reconstruction kernel on radiomic features leading to poor reproducibility. In a first step towards evaluating the ability of radiomic standardization techniques to provide robust quantitative features, the purpose of this investigation is to characterize the effect of CT reconstruction kernel on radiomic features in a DLD patient population, specifically on regions identified as normal and as fibrotic.

2. METHODS

The raw CT projection data of patients receiving a high resolution CT (HRCT) exam for DLD was collected from 3 CT scanners (Definition AS, Flash, and Force; Siemens Healthineers, Forchheim, Germany) and retrospectively reconstructed with 5 reconstruction kernels: smooth, medium-smooth, medium, medium-sharp, and sharp, corresponding to B30f, B40f, B45f, B60f, and B70f for Definition/Flash; Br32d, Br40d, Br49d, Br59d, Br69d for Force. The B45f and Br49d kernels are part of the clinical protocol at our institution and were considered reference kernels for this investigation. The non-reference kernels were chosen to sample kernels smoother and sharper relative to the reference kernel. All reconstructions were generated using Siemens ReconCT (version 14).

Regions of classic normal and classic fibrosis were annotated by 2 thoracic radiologists on images reconstructed with reference kernels (either B45f or Br49d). Annotations were pre-processed with a morphological erosion using a 7×7 structuring element to remove stray marks into areas of tissue other than normal or fibrotic that may affect radiomic feature values. Annotations were copied across each reconstruction, and Pyradiomics (version 3.1) was used to extract 72 radiomic features from each annotation across reconstruction conditions, including 18 first-order, 22 gray-level co-occurrence matrix (GLCM), 16 gray-level run-length matrix (GLRLM), and 16 gray-level size zone matrix (GLSZM) features.

To characterize the effect of CT reconstruction kernel on radiomic features, we examined the variability of features across kernels and assessed agreement between features in reference and non-reference kernels with concordance correlation coefficient (CCC). CCC across conditions and feature classes for both tissue phenotypes was visualized with heatmaps. A feature was considered robust for a given condition if CCC was > 0.9, and the total number of robust features was tallied for each feature class.

3. RESULTS

There were 66 patients included, with 116 areas of normal tissue and 208 areas of fibrosis annotated across all patients. Example image patches reconstructed with the B45f kernel demonstrating annotations of normal and fibrotic lung regions are shown in Fig. 1. Figure 2 displays example image patches across all reconstruction kernels for normal and fibrotic tissue. Smoother kernels reduce visual noise, but spatial resolution decreases. Conversely, spatial resolution increases in sharpness; however, so does visual noise.

Figure 1.

Example patches of normal and fibrotic tissue annotated by a thoracic radiologist on images reconstructed with the B45f kernel.

Figure 2.

Examples of normal (top) and fibrotic (bottom) tissue identified by a thoracic radiologist reconstructed with a variety of kernels ranging from smooth (B30f/Br32d) to sharp (B70f/Br69d). The appearance of finer texture details become visible as kernel sharpness increases; noise also increases as kernel sharpness increases.

Figure 3 displays the average value across patients and kernels for a stable (first-order root mean squared) and unstable (GLCM inverse variance) radiomic feature for both tissue types. In both normal and fibrotic tissue, reconstruction kernel had a minimal impact on first-order root mean squared, although the magnitude of values differed between tissue types. Alternatively, GLCM inverse variance generally decreased in normal tissue as kernel progresses from smoothest to sharpest. However, in normal tissue, the smoothest kernels yield slightly lower values compared to reference, and feature value quickly decreases as kernel sharpness increases.

Figure 3.

Line graphs of average value across patients for each reconstruction kernel for a radiomic feature with little variability (first-order root mean squared, left) and a feature with large variability (GLCM inverse variance, right). Note: Because 2 different radiomic features are shown, the y-axis scales do not correspond and should not be compared.

Heatmaps of CCC between features extracted from normal and fibrotic regions in images reconstructed with non-reference kernels and the reference kernel are shown in Fig. 4. Across kernels in normal tissue, 0/16 GLSZM, 1/16 GLRLM, 0/22 GLCM, and 5/18 first-order features were robust (average CCC > 0.9). Across kernels in fibrotic tissue, 1/16 GLSZM, 3/16 GLRLM, 3/22 GLCM, and 7/18 first-order features were robust. The effect magnitude is greater for sharper kernels compared to normal tissue (lower CCC). The effect of kernel on normal tissue is similar for kernels sharper than reference regardless of scanner. For fibrotic tissue, sharper Br kernels (from the Force scanner) have less impact on CCC than sharper B kernels (from the Flash/Definition scanners), with B70f decreasing CCC the most. The Br32d kernel has a similar impact on CCC as sharper kernels do for both tissue types.

Figure 4.

Heatmaps of CCC values between features extracted from normal (left) and fibrotic regions (right) across kernels relative to the reference kernel B45f or Br49d). Kernels are ordered smoothest to sharpest (for all scanners). Gray cells indicate features that created a zero division.

4. DISCUSSION

This study demonstrated the impact of CT reconstruction kernel on radiomic features extracted from normal and fibrotic tissue in patients with diffuse lung disease. Relative to a reference reconstruction kernel, smoother kernels decrease visual noise and spatial resolution, obscuring finer details, especially in fibrotic tissue. Alternatively, sharper kernels increase spatial resolution, enabling better visualization of finer details but at the expense of increased visual noise. The visual impact of smoother and sharper kernels is similar regardless of tissue type of scanner.

Although kernel affects image appearance consistently across tissue types and scanners, the impact on radiomic features is more variable; there are some radiomic features that are invariant to changes in reconstruction kernel and some that are very sensitive. Figure 3 depicts two such features, first-order root mean squared (RMS) and GLCM inverse variance. For RMS, there is clear separation between average value in normal and fibrotic tissue, but intra-tissue variability across kernels is very small, especially for normal tissue. Average value increases as kernel sharpness increases in fibrotic tissue, however the magnitude is relatively small and consistent with more subtle trends in normal tissue. Behavior of this feature is similar for all 3 scanners regardless of tissue type.

The response to different kernels in GLCM inverse mean is more variable. The average value in normal tissue is invariant to kernels smoother than reference for Definition/Flash scanners, but it sharply decreases for both sharper kernels. This pattern is exists for data from the Force scanner as well. In fibrotic tissue, the effect of kernel is more linear across scanners, with average inverse variance consistently decreasing as kernels go from smooth to sharp.

The CCC heatmaps in Fig. 4 indicate the level of agreement between features measured in the reference kernel and the non-reference kernels. On a whole, features from fibrotic tissue have better agreement with reference than from normal tissue, however within feature classes there is large disagreement. First order features generally have higher agreement than other feature classes regardless of tissue type (more lighter colored cells indicating higher CCC), but this only holds for smoother kernels (with the exception of Br32d). Nearly half of the first order features were robust to kernel (CCC > 0.9) for fibrotic tissue, and multiple GLCM and GLRLM features have higher agreement across kernels (indicated by lighter purple horizontally). There are only a few features per remaining class that agree well with reference in both tissues. The B40f and Br40d kernels yield the greatest agreement across features, however many GLCM features for these kernels do not agree well with reference. GLCM features generally agree with reference the least for fibrotic tissue, but improve in GLRLM and GLSZM. These 3 feature classes demonstrate similarly low agreement in normal tissue. The worst agreement over all features results from the B70f kernel regardless of tissue type.

This study contains some limitations. First, because reconstruction kernels do not perfectly match across scanners, there is some unavoidable inter-scanner variability in features. This potentially creates differences in values from reference kernels between scanners (Br49d for Force, B45f for Flash/Definition), making them not directly comparable to each other. We have tried to match corresponding kernels, however this difference is impossible to entirely mitigate. Second, the annotations of tissue phenotypes are not precise to tissue boundaries primarily because such boundaries are not always clearly defined in DLD. This results in some annotations potentially containing multiple tissue types. Some adjacent tissue types may be dramatically different from each in terms of attenuation and texture, and may therefore be affected differently by different reconstruction kernels. The morphological erosion of the annotations likely did not completely prevent multiple tissue types from being included, and this no doubt would affect radiomic feature values.

The findings of this investigation are consistent with other works investigating the impact of reconstruction kernel on radiomic features,⁷ and are likely not exclusive to DLD.^16,18 Future work will explore the impact of additional parameters such as CT dose and other DLD phenotypes such as ground glass opacity and honeycombing.

A major barrier to clinical adoption of quantitative features such as radiomics for quantitative characterization of disease pathologies is the variability in feature values as a result of image acquisition and reconstruction condition. One approach for addressing this is to standardize feature values across imaging conditions to a common baseline.¹⁹ The results of this investigation emphasize the need for reproducible quantitative image features, therefore, future work will also include evaluating the ability of applying physics-based radiomics standardization to mitigate the affects of acquisition and reconstruction parameters.

5. CONCLUSION

First order features were more robust than other features categories, and GLCM features were the least robust. The magnitude of impact of kernel is greater on radiomic features extracted from fibrotic tissue than for normal tissue. Concordance analysis showed that feature robustness depends on both feature class and CT reconstruction kernel.