Machine Learning-Based Evaluation of Post-Treatment Dosimetry for Predicting Response to Yttrium-90 Radioembolization In Hepatocellular Carcinoma
Abstract
Purpose
Radiation segmentectomy using yttrium-90 (Y‑90) transarterial radioembolization (TARE) is an established treatment for hepatocellular carcinoma (HCC), with the goal of achieving complete tumor necrosis through delivery of high doses (>400 Gy). Despite widespread clinical use, the prognostic value of post-treatment dosimetric parameters remains uncertain. This study evaluated the ability of dosimetric metrics derived from post-Y‑90 imaging to predict treatment response through a novel analysis.
Methods
A retrospective analysis was performed on 63 HCC tumors from 63 glass Y‑90 TARE treatments. Gross tumor volume (GTV) contours were delineated on diagnostic CT or MRI and transferred to post-treatment Y‑90 Bremsstrahlung SPECT/CT. Absorbed dose was calculated using the local deposition method. Random forest modeling was used to identify which dosimetric parameters were most strongly associated with imaging response. Response was dichotomized according to whether or not patients achieved complete response via modified RECIST criteria. Predictive performance was assessed using receiver operating characteristic (ROC) analysis with bootstrapping and cross-validation. The most informative metrics were further examined for potential prognostic trends.
Results
Key predictive parameters included specific activity (Bq/sphere), mean GTV dose, and dose to 98% of the GTV (D98). Models incorporating these parameters achieved a mean area under the ROC curve of 0.736. Among non-responders, 19 of 21 tumors (~90%) exhibited specific activity >800 Bq/sphere, compared with 47% of responders. On average, responders received higher mean GTV dose (659.5 Gy vs. 524.6 Gy) and higher D98 (220.2 Gy vs. 153.4 Gy), than non-responders.
Conclusion
Select dosimetric parameters demonstrate modest predictive value for imaging response following Y‑90 TARE for HCC. In cases treated with standard doses, lower specific activity per sphere was associated with better response, suggesting that dose distribution within the tumor may be more prognostically relevant than absorbed dose alone and that microdosimetry may play a larger role in determining response.