Adaptive statistical iterative reconstruction (ASIR) affects CT radiomics quantification in primary colorectal cancer

Objectives To investigate whether adaptive statistical iterative reconstruction (ASIR), a hybrid iterative CT image reconstruction algorithm, affects radiomics feature quantification in primary colorectal cancer compared to filtered back projection. Additionally, to establish whether radiomics from single-slice analysis undergo greater change than those from multi-slice analysis. Methods Following review board approval, contrast-enhanced CT studies from 32 prospective primary colorectal cancer patients were reconstructed with 20% ASIR level increments, from 0 to 100%. Radiomics analysis was applied to single-slice and multi-slice regions of interest outlining the tumour: 70 features, including statistical (first-, second- and high-order) and fractal radiomics, were generated per dataset. The effect of ASIR was calculated by means of multilevel linear regression. Results Twenty-eight CT datasets were suitable for analysis. Incremental ASIR levels determined a significant change (p < 0.001) in most statistical radiomics features, best described by a simple linear relationship. First-order statistical features, including mean, standard deviation, skewness, kurtosis, energy and entropy, underwent a relatively small change in both single-slice and multi-slice analysis (median standardised effect size B = 0.08). Second-order statistical features, including grey-level co-occurrence and difference matrices, underwent a greater change in single-slice analysis (median B = 0.36) than in multi-slice analysis (median B = 0.13). Fractal features underwent a significant change only in single-slice analysis (median B = 0.49). Conclusions Incremental levels of ASIR affect significantly CT radiomics quantification in primary colorectal cancer. Second-order statistical and fractal features derived from single-slice analysis undergo greater change than those from multi-slice analysis. Key Points • Incremental levels of ASIR determine a significant change in most statistical (first-, second- and high-order) CT radiomics features measured in primary colorectal cancer, best described by a linear relationship. • First-order statistical features undergo a small change, both from single-slice and multi-slice radiomics analyses. • Most second-order statistical features undergo a greater change in single-slice analysis than in multi-slice analysis. Fractal features are only affected in single-slice analysis.

Adaptive statistical iterative reconstruction CT Computed tomography FBP Filtered back projection GLCM Grey-level co-occurrence matrix GLDM Grey-level difference matrix GLRL Grey-level run length GLZSM Grey-level zone-size matrix MBIR Model-based iterative reconstruction NGTDM Neighborhood grey-tone difference matrix ROI Region of interest SD Standard deviation
To date, the majority of published retrospective CT radiomics studies have been based on images reconstructed with FBP alone. Given that iterative reconstruction has been adopted by all manufacturers in current CT scanners, the impact of iterative reconstruction algorithms on quantitative image features is highly relevant to the field of radiomics. We hypothesised that ASIR would alter feature values significantly compared with FBP, and alter features progressively with incremental ASIR weightings (percentage increments). Thus, the primary aim of our study was to investigate whether hybrid iterative reconstruction, specifically ASIR applied with incremental weightings, affects the quantification of radiomics features, including first-, second-and high-order statistical as well as fractal parameters, using primary colorectal cancer at peak contrast enhancement as an exemplar in light of promising data [29]. Our secondary aim was to establish whether features calculated from single-slice analysis (2-dimensional [2D] radiomics) are influenced to a greater degree than those from multi-slice analysis (3-dimensional [3D] radiomics).

Participants
Following institutional review board approval and informed consent, 32 consecutive patients with primary colorectal cancer underwent contrast-enhanced CT from January 2012 to July 2014 from a single institution (prospective trial, ISCTRN 95037515). Exclusion criteria were tumour diameter < 2 cm (to assure a sufficient number of CT voxels for analysis), impaired renal function (estimated glomerular filtration rate < 50 mL/min) and previous iodinated contrast media allergic reaction precluding administration of an iodinated contrast agent.

CT acquisition and reconstruction
CT was performed on a single Discovery 750 HD multidetector CT scanner (GE Healthcare). As part of a prospective research protocol, a dynamic contrast-enhanced CT centred on the primary cancer was acquired using the following protocol: 100 kV; 75 mAs; z-axis coverage, 4 cm; scan field of view, 50 cm; matrix, 512 × 512 mm; B30 soft reconstruction kernel; 5 mm reconstructed slice thickness; axial mode with 35 acquisition time points at a 1.5-s temporal resolution for 45 s and a 5-s temporal resolution thereafter for 120 s. To minimise bowel peristaltic movement, 20 mg of the spasmolytic agent hyoscine butylbromide (Buscopan; Boehringer Ingelheim) was administered intravenously prior to data acquisition unless contraindicated. The contrast agent was administered as follows: 50 mL of 370 mg/mL iodinated contrast agent (Niopam, Bracco) via a pump injector (Medrad Stellant dual syringe, Bayer Healthcare) at a rate of 5 mL/s, followed by a 50 ml saline chaser at the same rate. Mean CTDI Vol and DLP were 137.8 ± 15.3 mGy and 551.0 ± 61.2 mGy cm, respectively. The dynamic acquisition was reconstructed at the scanner with six different ASIR percentages: 0%, (equivalent to FBP) 20%, 40%, 60%, 80% and 100%, resulting in six separate datasets per patient (Fig. 1).

Image analysis
Image analysis was carried out by two readers in consensus (a clinical oncologist and a radiologist with 5 and 10 years CT experience, respectively) using the CT acquisition corresponding to peak tumour enhancement, in order to maximise the tumour contrast-to-noise ratio. A free-hand region of interest (ROI) was drawn around the tumour on 0% ASIR reconstructions to generate two datasets per patient: (1) a single axial image corresponding to the largest tumour area; (2) five contiguous axial images including tumour. ROIs were then copied onto 20-100% ASIR reconstructions to ensure these were identical.
Radiomics features were extracted using an in-house software based on Matlab (Mathworks), previously validated on a digital phantom developed as part of the Image Biomarker Standardisation Initiative [30]. A medium smoothing filter and 32 bin width were applied to the native DICOM images.

Statistical analysis
All radiomics features were regarded as continuous in nature. To allow for the fact that measurements from the same subject were likely to be more similar than from different subjects, the analysis was performed using multilevel linear regression. Two-level models were used, with individual measurements nested within patients. The shape of the relationship between ASIR levels and radiomics parameter values was examined. Initially, cubic, squared and linear terms for ASIR were included in the analysis. If the higher order terms (i.e. cubic and squared terms) were not found to be statistically significant, they were omitted and a simple linear relationship between variables was assumed. The distributions of the majority of radiomics parameters were such that the assumptions of the statistical methods were met. However, one parameter (fractal dimension lacunarity) had a particularly positively skewed distribution, and was thus analysed on the log scale. Regression coefficients along with corresponding confidence intervals were reported, representing the change in the CT parameter value for a 20% increase in ASIR. Standardised effect sizes (B) were calculated by the absolute regression coefficient divided by the betweensubject standard deviation for a 20-unit increase in ASIR. Conventionally, B values ≤ 0.3 are considered representative of a small effect; B values > 0.3 and < 0.8 are considered a moderate effect; ≥ 0.8, a large effect. All analyses were undertaken by a statistician using Stata, version 13.1 (StataCorp LP). A p value < 0.05 was taken to represent statistical significance.

Participants
No tumour was identified on CT in 1 participant (confirmed histological diagnosis of colorectal adenocarcinoma based on a resected polyp). Voxel spatial mismatch between 0% ASIR and subsequent ASIR reconstructed series precluded the analysis in 3 further participants. Therefore, 28 CT datasets were suitable for single-slice (2D) radiomics analysis. Multi-slice (3D) analysis was only possible in a subset of 23 participants, as 5 tumours were not sufficiently large or appropriately oriented to allow segmentation on at least 5 consecutive axial slices. Tumours had a mean diameter of 5.7 ± 1.74 cm and were located as follows: caecum, n = 4; ascending colon, n = 2; sigmoid colon, n = 7; rectum, n = 15. Radiological tumour (T) and nodal (N) staging, evaluated on CT images and based on the AJCC/UICC TNM classification of malignant tumours (8th edition), was as follows: T2, n = 9; T3, n = 17; T4, n = 2; N0, n = 16; N1, n = 10; N2, n = 2.

Radiomics analysis
Absolute regression coefficients, standardised effect size B values and corresponding p values for first-order, second-order and fractal radiomics features from single-slice and multi-slice datasets are summarised in Table 2; values for high-order features from multi-slice datasets are summarised in Table 3.

Single-slice analysis
All first-order, second-order and fractal features varied significantly and according to a linear relationship with increasing ASIR values, with the exception of GLCM sum entropy ( Table 2). Some features significantly increased; others decreased.
The relative effect was small on all first-order features (median standardised effect size B = 0.08; range, 0.02-0.24). Most second-order features, including all grey-level difference matrix (GLDM) and the majority of grey-level co-occurrence matrix (GLCM) features, were moderately affected (median B = 0.36; Describe the spatial interrelationship between neighbouring runs of voxels with the same intensity Run percentage, high grey-level run emphasis, short-run low grey-level emphasis, short-run high grey-level emphasis, short-run emphasis, long-run emphasis, grey-level non-uniformity, run length non-uniformity, low grey-level run emphasis, long-run low grey-level run emphasis, long-run high grey-level run emphasis, intensity variability, run length variability High-order grey-level zone-size matrix (GLZSM) Describe signal intensity and spatial interrelationship between neighbouring zones with the same intensity Short-zone emphasis, short-zone low-intensity emphasis, short-zone high-intensity emphasis, long-zone low-intensity emphasis, long-zone high-intensity emphasis, long-zone emphasis, intensity non-uniformity Model-based fractal features Describe repetitive patterns within an image extracted by using filter grids Fractal dimension mean, fractal dimension SD, fractal lacunarity, Hurst exponent, blanket mean, blanket max

Multi-slice analysis
Nearly all first-order, second-order and high-order features varied significantly with increasing ASIR values: some increasing, others decreasing. Exceptions were first-order range, neighbourhood grey-tone difference matrix (NGTDM) complexity and 3 out of 13 grey-level run length (GLRL) features (Tables 2 and 3). A small effect was confirmed on first-order statistical features (median B = 0.08; range, 0.01-0.27). Second-order features were affected to a lesser degree than from single-slice analysis (median B = 0.13; range, 0.07-0.35): 21 out of 23 features were affected in small measure, compared to 9 out of 23 from single-slice analysis. ASIR effect was small on high-order features (median B = 0.09; range, 0.01-0.31). In contrast to single-slice analysis and to all other features, multi-slice fractal features did not change significantly with increasing ASIR levels.

Discussion
Despite the rising number of studies investigating the clinical potential of radiomics in cancer imaging, relatively little is known  on how the shift of CT image reconstruction from filtered back projection to hybrid iterative algorithms might affect radiomics features quantitation. In our prospective study, we found that the application of incremental ASIR levels altered most statistical (first-, second-and high-order) features according to a linear relationship, both in single-slice and multi-slice analysis. Fractal features changed significantly only in single-slice datasets. While first-order features underwent small relative effects across all datasets, second-order features underwent a greater change in single-slice than in multi-slice analysis. A fundamental aspect of CT is the assignment of an attenuation value to a voxel. Analytical reconstruction algorithms, namely filtered back projection, have long been the backbone of CT reconstruction. However, as image noise from the Poisson statistical variation across an image is not accounted for by FBP; this has been a limiting factor for low-dose CT imaging [31]. Iterative reconstruction algorithms have thus risen to the fore with advances in scanner hardware and increasing computing power [31]. Hybrid algorithms such as ASIR combine both analytical and iterative methods, optimising image characteristics by decreasing image noise. However, there is a perceptive alteration in image characteristics with high ASIR increments often described as a 'waxy' appearance. Given this, we hypothesised that radiomics features may be altered by using ASIR compared to FBP and the ASIR weighting.
To date, only one other clinical study has compared the effect of different reconstruction algorithms on quantitative CT data including a limited number of histogram and second-order GLCM features. The effect of FBP, 50% ASIR and model-based iterative reconstruction (MBIR) was compared for volumes of interest in 20 patients with liver lesions (n = 13), lung nodules (n = 9) or renal calculi (n = 25) who underwent either non-enhanced CT or contrast-enhanced CT at a fixed 120 kVp but 2 different dose levels (full and halfdose) [32]. MBIR had the highest impact on features. Fifty percent ASIR had a significant effect on standard deviation (SD) but not on other first-order (entropy, kurtosis or skewness) or second-order GLCM features studied. These differences in part reflect a difference in study design, i.e. assessment of an incremental effect of ASIR as opposed to a comparison of a single ASIR percentage to FBP and MBIR.
We acknowledge that the reconstruction algorithm is only one of several factors that potentially affect CT radiomics features. Other factors include acquisition factors, e.g. kVp, mAs, reconstruction kernel, voxel size, grey-level discretisation and contrast administration [32][33][34][35][36][37][38][39]. For example, Zhao et al investigated the effect of reconstruction kernel on 89 unenhanced CT radiomics features including shape, first-order, second-order statistical, wavelet and fractal features for 32 lung cancers using 2D and 3D images. The reconstruction kernel had a significant effect on extracted features, with smooth images having a smaller effect than noisier images. 3D images were also more reproducible than 2D images [34]. This was also echoed in our study where 3D features were more stable.
He et al assessed the effects of reconstruction slice thickness, reconstruction kernel and contrast-enhancement on the diagnostic performance of 150 radiomics features in 240 patients with solitary pulmonary nodules (malignant, n = 180; benign, n = 60). This study demonstrated better discrimination and classification for malignant versus benign nodules when based on unenhanced versus contrast-enhanced CT, thin-(1.25 mm) versus thick-slice CT (5 mm) and standard versus lung reconstruction kernel [35]. In a further study, Shafiq-Ul-Hassan et al found that voxel size and discretisation were also important factors affecting radiomics features including shape, intensity, GLCM, GLZSM, GLRL, NGTDM, fractal and wavelet features in a digital phantom [39].
From our study and published data, minimisation of variation in reconstruction kernel, reconstruction algorithm, voxel size and grey-level discretisation would improve the quantification and stability of CT radiomics features.
We acknowledge some limitations to our study. Firstly, the study cohort was small (n = 28) but with the advantage of fixed acquisition parameters, including kVp and mAs, reconstruction kernel, and voxel size, allowing us to focus on the effect of the reconstruction algorithm alone. Secondly, we did not assess the effect of other iterative reconstruction algorithms as these were not available to us at the time of acquisition; we acknowledge that this would be of value going forward in future studies.
In conclusion, we confirmed that the application of a hybrid reconstruction algorithm versus traditional FBP affects CT radiomics quantification. The use of multi-slice (3D) rather than single-slice (2D) data will minimise this effect, particularly for second-order statistical and model-based fractal features. The reconstruction algorithm should be taken into account and standardised when acquiring data for future multicentre CT radiomics studies.

Compliance with ethical standards
Guarantor The scientific guarantor of this publication is Professor Vicky Goh.