Selective Internal Radiation Therapy With Yttrium-90 Glass Microspheres: Biases and Uncertainties in Absorbed Dose Calculations Between Clinical Dosimetry Models

Abstract

Purpose: To quantify differences that exist between dosimetry models used for ⁹⁰Y selective internal radiation therapy (SIRT).

Methods and materials: Retrospectively, 37 tumors were delineated on 19 post-therapy quantitative ⁹⁰Y single photon emission computed tomography/computed tomography scans. Using matched volumes of interest (VOIs), absorbed doses were reported using 3 dosimetry models: glass microsphere package insert standard model (SM), partition model (PM), and Monte Carlo (MC). Univariate linear regressions were performed to predict mean MC from SM and PM. Analysis was performed for 2 subsets: cases with a single tumor delineated (best case for PM), and cases with multiple tumors delineated (typical clinical scenario). Variability in PM from the ad hoc placement of a single spherical VOI to estimate the entire normal liver activity concentration for tumor (T) to nontumoral liver (NL) ratios (TNR) was investigated. We interpreted the slope of the resulting regression as bias and the 95% prediction interval (95%PI) as uncertainty. MC_NL^single represents MC absorbed doses to the NL for the single tumor patient subset; other combinations of calculations follow a similar naming convention.

Results: SM was unable to predict MC_T^single or MC_T^multiple (p>.12, 95%PI >±177 Gy). However, SM^single was able to predict (p<.012) MC_NL^single, albeit with large uncertainties; SM^single and SM^multiple yielded biases of 0.62 and 0.71, and 95%PI of ±40 and ± 32 Gy, respectively. PM_T^single and PM_T^multiple predicted (p<2E-6) MC_T^single and MC_T^multiple with biases of 0.52 and 0.54, and 95%PI of ±38 and ± 111 Gy, respectively. The TNR variability in PM_T^single increased the 95%PI for predicting MC_T^single (bias = 0.46 and 95%PI = ±103 Gy). The TNR variability in PM_T^multiple modified the bias when predicting MC_T^multiple (bias = 0.32 and 95%PI = ±110 Gy).

Conclusions: The SM is unable to predict mean MC tumor absorbed dose. The PM is statistically correlated with mean MC, but the resulting uncertainties in predicted MC are large. Large differences observed between dosimetry models for ⁹⁰Y SIRT warrant caution when interpreting published SIRT absorbed doses. To reduce uncertainty, we suggest the entire NL VOI be used for TNR estimates when using PM.

Figure 1

Spatial representation of the dosimetry models. Diagnostic CT (a), 90Y SPECT (b), SM (c), PM (d), MC (e). Gold, blue, red, and cyan color washes represent 20, 40, 55, and 130 Gy, respectively.

Figure 2

Illustration of the ideal, best-expected, and realistic case for PM calculations (a). TNR variability was introduced into PM absorbed doses by sampling the non-tumoral liver activity concentration with a single sphere multiple times (b). In total, the variability was characterized by 4 TNRs: using the entire non-tumoral liver VOI and 3 spheres. Example calculations in the figure were performed assuming M_NT =1.8 kg, M_T =0.2 kg, A=4.8 GBq, and a conversion constant of 50 Gy-kg/GBq

Figure 3

Box plots summarizing the absorbed doses from different dosimetry models over all patients for tumors (a) and non-tumoral liver (b). Default settings in R were used to calculate boxplot parameters. Red bars represent the median value. Top and bottom of boxes represent the 75^th and 25^th percentiles. Outliers are defined as a point greater (or less) than 1.5 times the interquartile range above (or below) the upper (or lower) quartile.

Figure 4

Predicting mean MC absorbed doses through linear regression of SM (left column), PM (center column), or PM+TNR variability (right column) for tumor (top two rows) and normal liver absorbed doses (bottom two rows). Both the single tumor subset (first and third rows) and multiple tumor subsets (second and fourth rows) are shown. Shaded bands represent 95% confidence intervals. Dashed lines represent 95% prediction intervals.

Figure 5

Summary of linear regression slopes (bias) and 95% prediction intervals (uncertainty) for SM, PM, and PM+TNR variability. Error bars represent the 95% confidence interval. SM was not displayed for linear regression of tumor absorbed doses because no statistically significant correlations were found between SM and MC for tumors.