Evaluation of AI-Driven Automatic Segmentation Algorithms for Lung Shunt Fraction Estimation in <sup>90</sup>Y-Labeled Microsphere Therapy.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, University of Texas Southwestern Medical Center, Dallas, Texas.
- Department of Radiology, University of Texas Southwestern Medical Center, Dallas, Texas [email protected].
Abstract
<sup>90</sup>Y-labeled microsphere (<sup>90</sup>Y-microsphere) hepatic arterial radioembolization is a direct treatment for hepatic neoplasms. Before therapy, the fraction of <sup>90</sup>Y-microspheres that may shunt past the liver to the lungs, or lung shunt fraction (LSF), is estimated using <sup>99m</sup>Tc-macroaggregated albumin. LSF is commonly estimated using regions of interest around the lungs and liver on planar images, a method limited by indeterminate margins and overlap between the lungs and liver. The estimation of LSF using SPECT/CT with manual segmentation is time-consuming and poorly reproducible. In this study, we evaluated 2 commercially available artificial intelligence (AI)-driven automatic segmentation algorithms for estimating pretreatment LSF in <sup>90</sup>Y-microsphere therapy. <b>Methods:</b> A retrospective analysis of 91 SPECT/CT studies from 43 patients who have underwent pretherapy <sup>99m</sup>Tc-macroaggregated albumin and posttherapy <sup>90</sup>Y SPECT/CT imaging was performed using 2 organ segmentation tools: AI-Rad Companion Organs RT (ORT) and Contour ProtégéAI+ (CPAI). All CTs were reconstructed using a B30s kernel, whereas an additional B60s kernel was included for 61 examinations. Manual adjustments were applied to the contours and SPECT data when needed (e.g., diaphragmatic misregistration). Organ volumes and LSFs were compared between the 2 AI tools and 2 reconstruction kernels. The repeatability of LSF and lung and liver volume measurements was tested on 20 CT studies. <b>Results:</b> ORT estimates of lung and liver volumes were significantly larger with the B60s kernel compared with the B30s kernel (lung, <i>P</i> < 0.001; liver, <i>P</i> = 0.006). Conversely, with CPAI, lung and liver volumes were significantly greater using the B30s kernel compared with the B60s kernel (lung, <i>P</i> = 0.010; liver, <i>P</i> = 0.010). Lung volumes were higher using CPAI versus ORT for both the B30s (<i>P</i> < 0.001) and B60s (<i>P</i> < 0.001) kernels. Consequently, LSFs calculated using CPAI were significantly higher than ORT (<i>P</i> = 0.002). Both AI tools exhibited excellent reproducibility, with an average percent coefficient of variation near 1%. <b>Conclusion:</b> AI software can produce highly repeatable organ segmentation contours, enabling more precise calculation of LSF. SPECT/CT-based LSF calculation may offer more accurate risk assessment and dose planning before <sup>90</sup>Y radioembolization compared with planar LSF estimation, although further validation is warranted before widespread adoption.