Verification of a method to detect glass microspheres via micro-CT.

ABSTRACT

INTRODUCTION: The ability to determine the microscopic distribution of glass microspheres in 90 Y radioembolization has important applications in post-treatment microdosimetry and cluster analysis. Current methods are time-intensive and labor-intensive and thus are typically only applied to small samples. MATERIALS AND METHODS: A high-resolution micro-CT image with a voxel size of 8.74 µm was acquired of phantoms containing ~25 µm-diameter glass microspheres embedded in tissue-equivalent materials that were optically transparent, which allowed true microsphere locations to be determined using transmission light microscopy. A 3-stage algorithm was developed to estimate the number and locations of microspheres in tissue regions. The stages are thresholding the CT image and discarding regions with insufficient voxels, estimating the number of microspheres in each region using the values of the detected and neighboring region voxels and estimating locations for each microsphere using the outputs of the previous two stages. Two different methods for estimating the number of microspheres in each region were derived, as were five methods for localizing microspheres. Metrics for each stage were computed, and the mean absolute error (MAE) between the dose to 72 µm voxels of the true and estimated dose maps created from the microsphere locations was used as the figure of merit for overall algorithm performance. Microsphere locations identified in the optical micrograph were used as the gold standard for the metrics of all stages. The method's utility was then demonstrated using a specimen from a human neuroendocrine tumor (NET) treated with glass 90 Y microspheres. RESULTS: The stage detecting regions containing microspheres found 100% of microspheres inside regions. The number of incorrectly detected regions without microspheres was 1.5% of the total number of regions. In stage 2, with these parameters, nearly 94% of the actual number of spheres in each region was correctly counted, and only 5% of the estimated sphere quantities in each region were false positives. The MAE between the true dose maps and dose maps estimated using the full algorithm with optimal parameter and method choices was 4.2%. A total of 5,713 glass microspheres were identified as being distributed heterogeneously in the NET specimen with a maximum tumor dose of >2500 Gy and 46% of the specimen receiving <20 Gy. CONCLUSIONS: This work developed and evaluated a method to detect and estimate the three-dimensional locations of glass microspheres in whole tissue samples that require less manual effort than traditional methods. This method could be used to gain important insights into the heterogeneity of microsphere distributions that would be useful for improving radioembolization treatment planning.