Skip to main content
SIR 2025 Event Main banner
BSCI - Consideration Cases Ad 1
Merit/DPSSME Siege
Back

SIR 2025

Interventional Oncology

Traditional Poster

Session: Traditional Poster Presentations

44 - Deep learning segmentation of cryoablation ice balls on intraprocedural imaging for automatic ablation zone quantification for soft tissue and bone ablations

Monday, March 31, 2025
12:10 PM – 1:10 PM CT
Location: Music City Center, Expo Poster Area
Purpose: The local control efficacy of any ablation modality is predicated on adequate ablation margins around the target lesion. A unique feature of cryoablation is the ability to visualize the ice ball during the ablation procedure. However, manual segmentation of ice balls during an ablation procedure to calculate ablation margins in real time is impractical. The purpose of this study was to develop an automatic ice ball segmentation algorithm utilizing a deep learning framework for bone and soft tissue cryoablation procedures. 

Materials and Methods: This retrospective study included 84 patients who underwent bone and soft tissue cryoablation between January 1, 2023 and June 1, 2024. Intraprocedural ice balls were manually segmented to serve as the ground truth. These patients were randomly distributed into training and validation cohorts across a 5-fold validation schema. A state-of-the-art deep learning framework (nnU-Net) was trained and validated utilizing the institutional computing cluster. Comparative metrics including Dice score were calculated against the manually annotated segmentations.  

Results: A total of 97 tumors were identified in the study. Among the patients, there were 34 females (40%) and 50 males (60%)(Table 1). The deep learning model demonstrated high accuracy, achieving a Dice score of 0.91. The model was successful across a diversity of tumor types. Furthermore, it was capable of defining ice ball margins even when the margin extended into tissues that obscure the ice ball directly such as adipose tissue or bone. The model was additionally effective at identifying ice balls across a range of size, ranging from 1cm to 10cm in long axis.  

Conclusion: A deep learning model successfully identifies intraprocedural ice balls during bone and soft tissue cryoablation. Automatic determination of ablation zones during cryoablation should facilitate quantitative assessment of ablation margins in real-time.