Potential of High Dimensional Radiomic Features to Assess Flowing Blood Components in Non-contrast CT Scans

Background: To assess the potential of radiomic features to quantify components of owing blood to non-invasively predict moderate-to-severe anemia in non-contrast enhanced CT scans. Methods: One hundred patients (median age, 69 years; range, 19–94 years) who received CT scans of the thoracolumbar spine and blood-testing for hemoglobin and hematocrit levels ± 24h between 08/2018 and 11/2019 were retrospectively included. Intraaortic blood was segmented using a spherical volume of interest with consecutive radiomic analysis. Feature selection was performed applying analysis of correlation and collinearity. The nal feature set was obtained to differentiate moderate-to-severe anemia. Random forest machine learning was applied and predictive performance was assessed. A decision-tree was obtained to propose a cut-off value of CT Hounseld units (HU). Results: High correlation with hemoglobin and hematocrit levels was shown for rst-order radiomic features (p<0.001 to p=0.032). The top 3 features showed high correlation to hemoglobin values (p) and minimal collinearity (r) to the top ranked feature Median (p<0.001), Energy (p=0.002, r=0.387), Minimum (p=0.032, r=0.437). Median (p<0.001) and Minimum (p=0.003) differed in moderate-to-severe anemia compared to non-anemic state. Median yielded superiority to the combination of Median and Minimum (p(AUC)=0.015, p(precision)=0.017, p(accuracy)=0.612) in the predictive performance employing random forest analysis. A Median HU value ≤ 36.5 indicated moderate-to-severe anemia (accuracy=0.90, precision=0.80). Conclusions: First-order radiomic features correlate with hemoglobin levels and may be feasible for the prediction of moderate-to-severe anemia. High dimensional radiomic features did not inherit the potential to augment the data in our exemplary use case of owing blood component assessment. Trial registration: Retrospectively registered. and hematocrit levels using high dimensional radiomic features in non-contrast enhanced CT scans. for collinearity (1). Features with a collinearity of r for further analyzed the obtained radiomic features set to differentiate moderate-to-severe anemic state. Moderate-to-severe anemia is dened by a cut-off value of hemoglobin ≤ 10–11 g/dL depending on age and gender (25–28). For our primarily methodologically driven study we aimed to choose a uniform denition of moderate-to-severe anemia and therefore dened a cut-off value of hemoglobin ≤ 10 g/dL for our cohort as previously proposed. We built two machine learning models based on random forest (RF) algorithms to predict moderate-to-severe anemia. The predictive power was assessed by receiver operating characteristics (ROC) with 100 fold cross-validation. Each run randomly drew 70% of the samples for training and tested the model with the remaining independent 30% of the We obtained the area under the curve (AUC), precision and accuracy. To analyze the variation of predictive power we applied a two-tailed student’s t test of the 100 fold cross-validated measurements. Machine learning algorithms and visualization of the decision tree were conducted in Python 3.7 using the open-source scikit-learn for one/(two) feature(s)) for RF prior normalization of features employing StandardScaler and DecisionTreeClassier with criterion = gini max_depth for the assessment of components of owing blood in our use case to assess hemoglobin and hematocrit levels. Based on our ndings, we conclude that higher dimensional radiomic features do not seem to be useful to predict components of owing structures, probably as potential diculties at each step of the radiomics workow may be more relevant in owing components. The application of radiomics may be limited to the assessment of solid tissues and tumor characteristics.

dimensional radiomic features did not inherit the potential to augment the data in our exemplary use case of owing blood component assessment.
Trial registration: Retrospectively registered.

Background
Radiomics is a term coined for computational quantitative imaging analysis and has been shown to inherit the potential in aiding clinical decision making (1). Radiomics extracts a large number of quantitative data from medical images that can provide surrogate information on biochemical and pathophysiological processes (2,3). The technique has been successfully applied to evaluate tumor characteristics non-invasively (4). While several studies showed the bene ts of radiomics in solid tissue and predominantly cancer research (5)(6)(7), its potential to assess owing structures and moving tissues has not yet been investigated.
Acute and chronic blood loss might not only be surrogates of yet undiagnosed diseased which require further workup but also might be considered as an illness itself which requires haemostasis management (8,9). In emergency patients with acute blood loss, fast assessment of a multitude of blood components, a.o. hemoglobin and hematocrit levels is essential (10,11). In 2002, the World Health Organization has attributed anemia as one of the most relevant risk factors leading to high mortality and morbidity (12,13). During hospitalization, phlebotomy is the current standard of screening for a load of blood components (14). Blood samples are usually easily obtained, but the procedure can be time consuming in some cases (15). Non-invasive screening of blood components in a clinically indicated CT may yield the potential to assess speci c blood components in order to focus invasive testing on pre-ltered components and patients to reduce workload and costs of laboratory analyses (16).
Computed tomography (CT) is a commonly used imaging modality in hospitalized patients and provides non-invasive assessment of tissue morphology.
Previous studies have suggested that simple attenuation measurements in CT scans correlate with hemoglobin and hematocrit levels and may be useful in predicting anemia (17)(18)(19).
By extracting a variety of mineable image features, radiomics can provide additional, higher dimensional data that can be employed to improve decision support. Current radiomic research promotes the impression that radiomic features are potentially applicable to augment data in a variety of diseases (1).
However, the potential of radiomic features to assess the owing blood compartment to predict speci c components has not yet been su ciently evaluated.
The aim of this study was to assess the predictability of hemoglobin and hematocrit levels using high dimensional radiomic features in non-contrast enhanced CT scans.

Patient selection
The local Ethics committee approved this retrospective study (project number: 20-689, Goethe University Frankfurt am Main, Germany) and waived informed written consent.
A total of 181 consecutive patients (female, 54; male, 46; age, 69 (19-94) years) who underwent non-contrast dual-energy CT imaging of the thoracolumbar spine between 08/2018 and 11/2019 were screened for study inclusion. Inclusion criteria were (I) > 18 years of age, (II) thoracolumbar region, (III) 1mm 90kV series, (IV) hemoglobin values ± 24h CT examination. Exclusion criteria were (I) different acquisition protocol, (II) signs of active bleeding, (III) imaging artifacts. All clinical data were obtained in clinical routine. 100 patients met the criteria and were evaluated. Figure 1 shows the owchart of patient inclusion.

Image reconstruction
We applied the 90kVp images as they are reconstructed using isotropic voxels in clinical routine (axial, section thickness 1mm and increment of 1mm) with a

Radiomic analysis
The 3D Slicer software platform (http://slicer.org, version 4.9.0) was applied to visualize and process the DICOM image stack (2,20). For segmentation, a radiologist (SM) with two years of experience manually de ned a spheric volume of interest (VOI, 1.0 cm diameter) centrically in the aorta of the thoracolumbar region, sparing the aortic wall and visual artifacts (Fig. 2). All VOIs were reviewed by a second radiologist (SB, two years of experience). Both radiologists were blinded to the laboratory results. Prior to feature extraction we did not perform further image manipulation as the Imaging Biomarker Standardization (IBSI) does currently not cover image preprocessing and we did perform our analysis on isotropic 1mm x 1mm voxels (21). The open-source package PyRadiomics was used as extension within 3D Slicer to extract the radiomic features (2,20,22 (24). We ranked the features according to the obtained p-value of the correlation analysis. The lower the p-value, the higher the ranking. Next, we used inter-correlation analysis of the features which showed signi cant correlation for both hemoglobin and hematocrit levels to test for collinearity (1). Features with a collinearity of r < 0.5 were selected for further analysis. Next, we analyzed the obtained radiomic features set to differentiate moderate-to-severe anemic state. Moderate-to-severe anemia is de ned by a cut-off value of hemoglobin ≤ 10-11 g/dL depending on age and gender (25)(26)(27)(28). For our primarily methodologically driven study we aimed to choose a uniform de nition of moderate-to-severe anemia and therefore de ned a cut-off value of hemoglobin ≤ 10 g/dL for our cohort as previously proposed. We built two machine learning models based on random forest (RF) algorithms to predict moderate-to-severe anemia. The predictive power was assessed by receiver operating characteristics (ROC) curves with 100 fold cross-validation. Each run randomly drew 70% of the samples for training and tested the model with the remaining independent 30% of the data. We obtained the area under the curve (AUC), precision and accuracy. To analyze the variation of predictive power we applied a two-tailed student's t test of the 100 fold cross-validated measurements. Machine learning algorithms and visualization of the decision tree were conducted in Python 3.7 using the open-source scikit-learn 0.21.3 packages RandomForestClassi er (n_estimators = 100, max_depth = 1/(2) for one/(two) feature(s)) for RF analysis with prior normalization of features employing StandardScaler (https://scikit-learn.org/) and DecisionTreeClassi er with criterion = gini and max_depth equivalent to the RF-analysis (29). Further statistical analyses were performed using Prism 6.0 (GraphPad software) and JMP 14 (SAS, Cary, U.S.A.). The signi cance values were indicated as followed: * p < 0.05; ** p < 0.01; *** p < 0.001. The respective table and gure legends give detailed information about the statistical tests.

Results
From all radiomic features, 9 features revealed signi cant correlation (p < 0.001-p = 0.032) to hemoglobin and hematocrit levels with Median (p < 0.001) as the highest ranked feature ( Table 2). The features were found to be part of one feature class, the rst-order statistics ( Table 2). Grey Level Non Uniformity, a feature of the GLSZM feature class, showed correlation to hematocrit levels, but no signi cance to hemoglobin levels ( Table 2). It was therefore excluded for further analysis. The selected features showed a high degree of collinearity (Fig. 3A, Table 3). Energy (r = 0.387), Maximum (r = 0.411) and Minimum (r = 0.437) were found to be the least correlated features to Median ( Table 3). As Maximum revealed collinearity with Energy (r = 0.568) it was excluded for further analysis. We therefore obtained the top 3 features to correlate with hemoglobin and hematocrit levels: Median (p < 0.001, Fig. 3B), Energy (p = 0.002, Fig. 3C) and Minimum (p = 0.032, Fig. 3D). Multivariate measurements of correlations of radiomic features that are signi cantly correlated with hemoglobin and hematocrit levels.
Radiomic analysis of intraaortic blood to differentiate a threshold of hemoglobin level of 10 mg/dL revealed signi cant difference in the radiomic features Median (p < 0.001, Fig. 4A) and Minimum (p = 0.003, Fig. 4B) whereas Energy did not reach the level of signi cance (p = 0.09, Fig. 4C) and was therefore excluded for the consecutive machine learning model development.
A random forest based, 100 fold cross-validated machine learning approach was conducted applying either Median and Minimum features (Fig. 5A, AUC 0.88 ± 0.07) or Median feature only (Fig. 5B, AUC 0.90 ± 0.06) for model building. Application of the single radiomic feature Median was superior to its combination with the feature Minimum with regard to AUC and precision measurements whereas no difference was found with regard to model accuracy (Fig. 5C, accuracy p = 0.612, AUC p = 0.015, precision p = 0.017).
We obtained a decision tree based on the radiomic feature Median (Fig. 5D). With a cutoff value of ≤ 36.5 Houns eld Units (HU) in an independent train/test set of patients drawn at random, we obtained a test accuracy of 0.90 and precision of 0.80 to predict moderate-to-severe anemic state.

Discussion
In this study, we examined the potential of high dimensional radiomic features to assess components of the moving blood compartment. We assumed that hemoglobin and hematocrit may be the most promising and easily non-invasively accessible values and may inherit the clinical potential to predict moderate-to-severe anemia. Examining 100 non-enhanced CT scans, we demonstrated correlation of rst-order radiomic features with hemoglobin and hematocrit levels. We could obtain a cut-off value of ≤ 36.5 HU for Median to predict moderate-to-severe anemia with an accuracy of 0.90 and a precision of 0.80. We could show that higher dimensional radiomic features did not augment simple rst order radiomic features. Based on our ndings, we conclude that besides its bene t to evaluate solid tissue and tumor characteristics non-invasively, the application of higher dimensional radiomic features to analyze owing structures such as the blood system does not seem to be promising.
Our results regarding rst order radiomic features are in accordance with previous studies investigating the potential of quantitative measurements of CT density to differentiate between anemic and non-anemic conditions (27,30,31). In a study of 102 patients undergoing thoracic CT scans, the authors obtained mean attenuation measurements in the left ventricle which performed better than subjective reviewer analysis (27). Another study revealed a correlation between mean attenuation values of the thoracic aorta and hemoglobin values (30). Nevertheless, these studies did not include higher dimensional radiomic features, limiting their quantitative assessment to the mean value only (30).
Quantitative imaging data have been increasingly applied in the last years. Especially in cancer research, radiomics is a rapidly evolving research eld (32,33). In contrast to results obtained from research of speci c tissues or tumor types, our data suggest that the application of high dimensional radiomic features may not yield diagnostic value assessing owing structures, such as speci c components of the blood stream. In our study, high dimensional radiomic features were inferior to simple rst order statistic values to estimate hemoglobin or hematocrit values and they were not applicable to predict moderate-to-severe anemia. However, rst-order histogram features did signi cantly correlate with hemoglobin and hematocrit values with promising predictive power of therapeutically relevant anemic state.
Potential problems at each step of the radiomics work ow including image acquisition, image reconstruction, segmentation and pre-processing have already been described in literature (34). In their study from 2020, Fornacon-Wood at al. argued that different acquisition protocols (35), image reconstruction algorithms, reconstruction parameters (kernel) (36) and number of grey levels used to discretize histogram (37) affect feature values and feature reproducibility. Our study suggests that these issues seem to be more relevant in moving and dynamic compartments as high dimensional radiomic features had no diagnostic power for the prediction of hemoglobin and hematocrit levels. This raises the question whether most of the measured texture in a noncontrast-enhanced CT blood pool may be the effect of imaging artifacts due to the laminar ow of the blood system rather than true data of biological components.
Our study has limitations that warrant discussion. Analyzing retrospective data with continuous patient enrollment, we cannot rule out a selection bias. We had a moderate bias towards females and the older population and cannot rule out that a more balanced study population might have altered the results. Depending on age and gender, moderate-to-severe anemia is de ned by a cut-off value of hemoglobin ≤ 10-11 g/dL (25-28). As previously described, we chose a uniform cut-off value of hemoglobin ≤ 10 g/dL for our primarily methodologically driven study but we cannot rule out that age, gender or pregnancy adjusted values might have altered the results. Our study design was limited to 100 patients and a bigger cohort might have been favorable. This bias might reduce generalizability of the results and the nally obtained cut-off value of 36.5 HU to differentiate moderate-to-severe anemic state. We restricted the patient inclusion to one dual-energy CT scanner to exclude inter-scanner variability and to include only reconstructions with 1mm isotropic voxels, nevertheless, intra-scanner variability may have occurred. We limited the region of VOI de nition to the thoracolumbar region to have an adequate diameter of the aorta for VOI placement and to limit pulsation artifacts that might be present at the ascending thoracic aorta.

Conclusions
CT is a commonly applied imaging modality for a multitude of diagnostic purposes and attenuation measurements of various degrees of complexity are easily performed. We obtained simple histogram and high dimensional radiomic features and could demonstrate that histogram radiomic features enable an accurate differentiation of moderate-to-severe anaemic state and non-anemic state employing non-enhanced CT scans. We emphasize that our results are the rst to show that high dimensional radiomic features are inferior to simple histogram features and do not yield additional information for the assessment of components of owing blood in our use case to assess hemoglobin and hematocrit levels. Based on our ndings, we conclude that higher dimensional radiomic features do not seem to be useful to predict components of owing structures, probably as potential di culties at each step of the radiomics work ow may be more relevant in owing components. The application of radiomics may be limited to the assessment of solid tissues and tumor characteristics. Availability of data and materials

Abbreviations
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests.

Funding
This study was not supported by any funding. Representative images of the measurement technique Axial (A), sagittal (B) and coronal (C) plane with 3D-volume rendering (D) of a standard volume of interest (VOI) placement is shown in a patient with a hemoglobin and hematocrit level of 7.2 g/dL and 22.4%, respectively. A spherical VOI with 1cm in diameter was placed within the lumen of the thoracoabdominal aorta as described in detail in the materials and methods section.  Median density measurement of Houns eld units reveals the best working model to predict moderate-to-severe anemia Analysis of prediction performance for moderate-to-severe anemia with 2 variant feature subsets applying random forest (RF) machine learning algorithms (A-C). 100 fold cross-validated (colors) receiver operating characteristics (ROC) curve analysis of the validation cohort with mean ROC curve (blue) and ± 1 standard deviation (grey area) are shown for Median and Minimum (A) or Median only (B). RF maximum depth was 2 (A) and 1 (B). C shows the Box-Whisker Plots with 5-95% percentile for both cross-validated prediction models with the respective accuracy, area under the curve (AUC) and precision. Two-tailed, unpaired student's t-test was applied for model comparison (C, p-values). In D, an exemplary decision tree with a depth of 1 for Median is shown to stratify moderate-to-severe anemic state (y0= Hb>10 g/dL; y1= Hb≤10 g/dL). The decision tree with Gini-based tting applied a training cohort of 70% drawn at random and an independent test cohort of 30%.