290 - Integrating AI and Deformable Registration into a Fully Automated Tool to Quantify the Ablation Margin for Renal Cell Carcinoma

Purpose: To develop and demonstrate the clinical utility of a fully-automatic AI segmentation and deformable registration tool to quantify achieved tumor treatment margins for renal cell carcinoma (RCC) ablation.

Materials and Methods: Pre- and post-ablation CT images from RCC patients treated at a single institution were utilized for training and validation of the tool-kit. Tumor and ablation zone segmentations using a U-Net based framework, trained on a manual segmentation dataset were automated on CT images collected for validation (n=46), and the deformable registration between corresponding pairs of pre- and post-ablation CT images was performed. The registration tool was used for complete surface evaluation between tumor and ablation zone to quantify the mean ablation margin achieved and was compared to rates of local RCC recurrence to assess clinical utility.

Results: Patient demographics are presented in Table 1. The mean Dice similarity coefficient (DSC) between the AI and manual segmentation was 0.86 ± 0.05. There was no statistical significance (p >0.05) in DSC when comparing ≥50% exophytic (0.85 ± 0.05), < 50% exophytic (0.85 ± 0.05) and endophytic (0.89 ± 0.05) tumors, and between radiofrequency ablation (0.86 ± 0.05) and cryoablation (0.85 ± 0.06).



17% of patients (8/46) had RCC recurrence occurring a median of 5.2 (3.8 - 6.8) years after successful ablation. All initial post-ablation CT images (taken 68 ± 49 days post-ablation) showed no viable tumor. However, the mean ablation margin calculated from these images using the deformable registration tool for patients who developed RCC recurrence was -0.2 ± 1.7 mm. This was significantly smaller (p< 0.001) than the ablation margin of those without RCC recurrence (3.5 ± 2.0 mm, n=38).

Conclusion: The developed fully-automatic tool demonstrates feasibility for combining AI for segmentation of the pre-ablation tumor and post-ablation zone with deformable kidney registration for RCC ablation margin quantification. The tool found a significant difference in the achieved mean ablation margin between patients with and without long-term RCC recurrence, suggesting that quantifying ablation margins may predict oncologic outcomes. The fully automated tool can be easily incorporated into routine clinical workflow.