Abstract
241514
Introduction: Yttrium-90 (Y90) microsphere hepatic arterial radioembolization is a direct ablative therapy for hepatic primary and metastatic neoplasms. Prior to therapy, the fraction of the microspheres that may shunt past the liver and potentially cause radiation-induced damage to the lungs is estimated using Technetium-99m macroaggregated albumin (MAA). Lung shunt fraction (LSF), the percentage of total counts within the lungs is often calculated by drawing region-of-interest (ROI) around the lungs and liver using planar images, Thisis challenging due to indeterminate margins and the overlap between the lungs and liver. Estimating LSF using SPECT/CT cross-sectional imaging is more accurate, though manual 3D organ segmentation is time-consuming and poorly reproducible. We evaluated the effectiveness of artificial intelligence (AI) organ segmentation software in accurate estimating LSF using SPECT/CT imaging.
Methods: A retrospective analysis of 79 SPECT/CT scans from 37 patients who have undergone pretherapy MAA and post-therapy Y90 SPECT/CT imaging was performed using two commercially available organ segmentation tools: Siemens Healthineers’ AI-Rad Companion Organs RT (ORT) and MIM Software’s Contour ProtégéAI (CPAI). All CTs were reconstructed in B30 and B08 kernels, while an additional B60 kernel was applied to 14 exams. The B08 kernel is utilized for attenuation correction in SPECT data. The structure sets containing liver and lung contours for each CT were exported to MIM software. Manual adjustments to the organ contours were made if errors were present, such as bowel gas included in lung contours or ascites fluid included in liver contours. SPECT files were realigned to correct for misregistration secondary to motion. Following corrections, the MIM SurePlan Liver Y90 workflow was used on the CT, SPECT, and structure set files to determine each study’s LSF, lung/liver counts, and lung/liver volume. Repeatability tests were done on 18 SPECT/CTs from 3 patients who did not require manual recontouring. Lung volume and LSF were only calculated from 14 patients whose entire lung fields were included in their pre and post therapy SPECT/CTs.
Results: The LSFs calculated using CPAI were significantly higher than the ORT LSFs, with an average percentage difference of 14.6% (P=0.01). Both lung and liver volumes from ORT showed a significant increase when utilizing the B60 kernel compared to B30 (lung p<0.001, liver p=0.006). Conversely, with CPAI, lung and liver volumes were both significantly greater using B30 compared to B60 (lung p=0.010, liver p=0.010). Lung volumes were higher when using CPAI compared to ORT for both B30 (p<0.001) and B60 (p<0.001) kernels. As a result, LSF was calculated consistently higher with CPAI than with ORT in every case. There was no significant difference in CPAI and ORT liver volumes for both B30 and B60. Both AI tools exhibited excellent reproducibility, with an average coefficient of variation of 0.6% for ORT and 1.6% for CPAI. All cases requiring manual recontouring were associated with CPAI segmentation.
Conclusions: AI software can rapidly produce highly accurate and repeatable organ segmentation contours, enabling a more precise calculation of LSF in comparison to planar calculation and manual segmentation. SPECT/CT-based lung shunt calculation therefore can provide more accurate risk assessment and dose calculation prior to Y90 radioembolization. Overall, ORT demonstrated greater robustness, requiring almost no manual adjustment, while CPAI required frequent manual corrections, exposing it to more potential for operator error. Lung shunt fractions from both tools were significantly different than planar lung shunt estimations, and LSFs were higher from CPAI due to consistently higher lung volume measurements. This study emphasizes the importance of conducting a comprehensive evaluation of AI software before incorporating it into clinical diagnosis or treatment planning.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.