Towards artificial intelligence based preoperative dosimetry for liver<sup>90</sup>Y selective internal radiation therapy.
Authors
Affiliations (5)
Affiliations (5)
- LTSI, Université de Rennes, Campus de Beaulieu Bât 22, Rennes, 35042, France.
- Therenva, 74f Rue de Paris, Rennes, 35000, France.
- Laboratoire du Traitement du Signal et de l'Image, Université de Rennes, Université de Rennes, Campus de Beaulieu Bât 22, Rennes, 35042, France.
- LTSI, 4893+VQ, Av. de la Bataille Flandres-Dunkerque CS 44229, Rennes, Brittany, 35000, France.
- LTSI, Université de Rennes - Campus Beaulieu, 263 Avenue du Général Leclerc, Rennes, Brittany, 35042, France.
Abstract
Selective Internal Radiation Therapy (SIRT) delivers 90 Y microspheres through the hepatic arterial system to achieve tumoricidal absorbed dose while limiting normal-tissue irradiation. Current pre-treatment dosimetry requires angiography and 99m Tc-MAA imaging, resulting in an invasive and resource-intensive workflow. To evaluate the feasibility of estimating post-therapy 90 Y dose distribution directly from diagnostic multiphasic CT using an automated deep-learning pipeline, without requiring angiography or surrogate-particle imaging.
Methods:In 152 treatments from the multicenter PROACTIF registry, arterial-and portal-phase CT, together with clinically documented perfusion territories and injected activities, were processed through automated segmentation, rigid registration, differential CT generation, and a generative model synthesizing a PET-like 90 Y activity map.Absorbed-dose maps were computed using the Local Dose Deposition method, and predictive accuracy was assessed using tumor dose-volume metrics and voxelwise spatial agreement.
Results: For the best-performing configuration, the median tumor D70 error was 23 Gy, with substantial inter-patient variability reflecting the challenges of CT-only dose prediction. High-dose coverage (V200) showed a similar pattern.In contrast, spatial agreement was consistently strong, with perfused-volume gamma passing rates around 93% and tumor-level rates around 85%, indicating that the macro-scale distribution of microspheres can be approximated from diagnostic CT. The full pipeline was fully automated and produced predictive dosimetry in under 30 seconds.
Conclusion:Pre-angiography CT contains exploitable information about arterial enhancement and perfusion patterns that enables feasible CT-only prediction of 90 Y spatial distribution. While tumor-level absorbed-dose metrics remain variable, the strong spatial agreement suggests potential utility as an early planning tool to support patient selection.