| Methodology |
-
Our study cohort consisted of n=1200 subjects drawn from a reference population of 3068 adult kidney donor candidates (5792 CT scans) performed as part of evaluation for kidney donation, which we have previously studied (Derstine et al. 2021). Patients were included if they were over 18 years of age; had CT scan performed with a GE Discovery or Lightspeed scanner, using GE Standard convolution kernel at 120 kVp, with the entire spleen volume visible in the scan extent; had height and weight recorded in their medical record at the time of evaluation; and were medically and psychosocially approved for donation. For each patient, the non-contrast phase image series was preferentially selected but arterial phase was selected if the non-contrast images did not meet inclusion criteria. A subset of 272 patients who had both a non-contrast and arterial phase CT scan from the same study instance were identified and used for paired comparisons.
-
All spleen segmentations were performed using a 3D deep learning convolutional neural network (CNN) with multi-scale attention and feature fusion modules, trained on 128x128x128 image volume and label volume pairs. Model training was performed iteratively, using 60 randomly chosen scans to build an initial spleen segmentation model. Segmentation results from this first model were then reviewed across all study scans, 20 scans with correct ground truth but inadequate model results were added to the training set, and the model was re-trained on 80 scans. This process was repeated 4 times, resulting in a total of 140 training scans, after which review of all scans showed correctly segmented spleens across the full study population. Review consisted of visual inspection of the predicted spleen boundaries overlaid on a collection of 9 equi-spaced transverse/coronal/sagittal slices (i.e., 27 total images across all 3 image planes) cropped to the spleen's bounding box.
-
Ground truth spleen borders were drawn by trained operators on 20 axial slices spanning the inferior to superior boundaries of each chosen spleen. To increase diversity in the training set, image augmentation was used. A series of five 128x128x128 voxel volumes (pairs of interpolated image data and label data) were created for each segmented spleen, cropped to boxes surrounding the spleen with randomly varied padding amounts between 5mm and 35mm. All slices containing spleen were used for model training; to avoid step artifacts, shape-based interpolation (Sven, Blended 3D Poly2Mask, 2013) was used to interpolate spleen boundaries on slice locations between the drawn ground truth slices. The resulting CNN was used to extract a 3D volumetric segmentation of each spleen.
-
3D spleen mesh volume in $cm^3$ (mL) along with seven commonly used linear spleen measurements were extracted via fully-automated programmatic scripts. Spleen width (W) was obtained and measured as "the maximum dimension of the spleen on axial images" (Joiner et al. 2015, Nuffer et al. 2017).Two different spleen height (a.k.a. length) measurements were obtained; the cranio-caudal distance between the inferior-most and superior-most points of the spleen (H1) (Yetter et al. 2003, Joiner et al. 2015, Nuffer et al. 2017), and "the maximum dimension of the spleen on coronal reformatted images" (H2) (Nuffer et al. 2017, Yetter et al. 2003). Four different spleen thickness measurements were obtained; the maximum splenic thickness (perpendicular to W) on the slice where W was determined (T1), the splenic thickness at the midpoint of W on the slice where W was determined (T2), the maximum thickness on any axial slice, perpendicular to the width on that slice (T3), and the splenic thickness at the midpoint of the width on the slice where T3 was determined (T4) (Bezerra et al. 2004, Nuffer et al. 2017).
-
Demographics, CT parameters, and spleen measurements were summarized separately for men and women, reporting mean and standard deviation (SD) for continuous variables and proportion for categorical variables. Means were compared using two-tailed t-tests assuming unequal variance, and proportions were compared using the $\chi$-squared test. Paired t-tests were used to compare the within subject differences between non-contrast and arterial phase spleen volumes in the paired subset. Pearson correlations and scatter plots were used to assess the linear relationship between CT measurements versus age, height, weight, and BMI.
-
Allometric regression models were used to find the optimal integer coefficient for the relationship between height and each spleen measurement (SV, H, W, and T). The allometric model $SV = \alpha \times height^{\beta} \times age^{\gamma}$ was transformed into the logarithmic form $log_e(SV) = \alpha + \beta log_e(height) + \gamma log_e(age) + \epsilon$ and linear regression was used to find the $\beta$ coefficient (optimal power of height) (Kaitaniemi, 2004). The resulting coefficient rounded to 3 (SV) or 1 (H, W, T) as the nearest integer for height, ergo height-adjusted spleen measures were computed as $SVI_{Ht} = SV/(height^3)$, $H_{Ht} = H/height)$, $W_{Ht} = W/height)$, and $T_{Ht} = T/height)$. Secondarily, an allometric regression model was used to find the optimal coefficient for the relationship between $SVI_{Ht}$ and age. The resulting coefficient was -0.21 which we rounded to -0.25 ($-1/4$) for ease of interpretation and practical application.
-
We calculated cut-points for mild-moderate splenomegaly as 2SD above the mean and massive splenomegaly as 5SD above the mean. Volume cut-points were rounded to the nearest integer, H/W/T cut-points were rounded to the nearest tenth. For linear measurements, we also calculated optimal cut-points for splenomegaly using the pROC package `coords' function in R, with the SV and $SVI_{Ht}$ cut-points as reference. Optimal cut-points for massive splenomegaly could not be calculated due to sample size limitations.
-
An alpha level of .01 was used to determine statistical significance. All statistical tests were performed in R version 4.3.2, using the package `ggplot2' for data visualization.
-
This study was approved by the Institutional Review Board of the University of Michigan with a waiver of informed consent. All data were originally collected during the course of standard medical treatment and were used retrospectively for analysis. Patient identifiers were removed and data coded during the study. The compiled dataset used for publication and uploaded to Deep Blue carries no identifiers and the individual records are not linkable to individual patients. All methods were performed in accordance with the relevant guidelines and regulations of the United States.
|
| Description |
-
Previously reported references ranges for spleen volume were developed in small cohorts, used estimates of spleen size rather than actual volumes, or did not adjust for person height. We aimed to develop healthy reference ranges for spleen volumes and linear measurements with appropriate height adjustment.
-
We used a machine learning model to extract spleen volumes (SV) from CT exams for 1200 healthy adult kidney donor candidates. We calculated sex-specific reference distributions for SV and seven common measures of spleen height (H), width (W), and thickness (T). We assessed Pearson correlations of each spleen measure versus age, height, weight, and BMI. We used allometric analysis to find the optimal height scaling powers for each measure. We proposed novel height-adjusted indices for each spleen measure and a height-and-age adjusted index for SV. We reported distributional splenomegaly (mean + 2SD) and massive splenomegaly (mean + 5SD) cut-points, and computed optimal cut-points for each H, W, and T measure using AUROC with the volume-based measures as reference.
-
SV was significantly, positively correlated with person height, weight, and BMI, and negatively correlated with age. Men have larger SV than women. SV divided by person height-cubed was the optimal height adjustment factor, so we calculated a height index; $SVI_{Ht} = SV [cm^3]/(height [m]^3)$. Cut-points for height-adjusted spleen volume were 74 (splenomegaly) and 122 (massive splenomegaly). $Age^(1/4)$ was the optimal age adjustment factor, so we calculated a height-and-age index; $SVI_{HtAge} = (SV [cm^3]/(height [m]^3)) * age [years]^(0.25)$. Cut-points for height-and-age adjusted spleen volume were 183 (splenomegaly) and 303 (massive splenomegaly). H, W, and T measures divided by person height were the optimal height adjustment factors, so we calculated height indices; $H [cm]/height [m]$, $W [cm]/height [m]$, and $T [cm]/height [m]$.
-
We extracted spleen volumes and linear measurements using high-throughput volumetric analyses and report healthy reference ranges and splenomegaly cut-points neutral to sex and adjusted for height and age.
|
|---|