Prediction of Frozen Region Growth in Kidney Cryoablation Intervention Using a 3D Flow-Matching Model

TL;DR

The paper addresses real-time prediction of iceball growth during kidney cryoablation by introducing a 3D flow-matching framework that learns a residual deformation field to map early CT images to future states. The method represents the transformation as $I(τ)=I_{src}+τ(I_{tgt}-I_{src})$ with a residual velocity $u(τ)=r$, estimated by a neural network and integrated via an ODE solver to produce both a future CT volume and iceball segmentation. Compared to a diffusion-based baseline, the Flow approach achieves higher segmentation accuracy and image fidelity, with IoU reaching $0.61$ and Dice $0.75$ at $100$ iterations, and enables ~15 seconds per inference on a single A100 GPU. This data-driven, patch-wise strategy has the potential to enhance intraoperative guidance and minimize injury to healthy tissue, while remaining computationally feasible for real-time clinical use; future work includes multicenter validation and incorporating additional physiological variables to further improve robustness and clinical utility.

Abstract

This study presents a 3D flow-matching model designed to predict the progression of the frozen region (iceball) during kidney cryoablation. Precise intraoperative guidance is critical in cryoablation to ensure complete tumor eradication while preserving adjacent healthy tissue. However, conventional methods, typically based on physics driven or diffusion based simulations, are computationally demanding and often struggle to represent complex anatomical structures accurately. To address these limitations, our approach leverages intraoperative CT imaging to inform the model. The proposed 3D flow matching model is trained to learn a continuous deformation field that maps early-stage CT scans to future predictions. This transformation not only estimates the volumetric expansion of the iceball but also generates corresponding segmentation masks, effectively capturing spatial and morphological changes over time. Quantitative analysis highlights the model robustness, demonstrating strong agreement between predictions and ground-truth segmentations. The model achieves an Intersection over Union (IoU) score of 0.61 and a Dice coefficient of 0.75. By integrating real time CT imaging with advanced deep learning techniques, this approach has the potential to enhance intraoperative guidance in kidney cryoablation, improving procedural outcomes and advancing the field of minimally invasive surgery.

Figures (3)

• Figure 1: (A) Data collection and labeling pipeline, showing semi-automated segmentation of 3D CT scans acquired at 3-10 minutes to generate corresponding masks. (B) Patch-wise model training and generation framework, employing a 3D U-Net–based flow-matching approach with iterative ODE sampling to produce time-specific CT volumes and segmentation masks.
• Figure 2: (A) Comparison of 3D surface models of the Iceball region, contrasting the ground truth (green) with the generated region (red) overlaid on axial CT slices. (B) Slice-by-slice comparison of the Iceball region masks, illustrating input and generated CTs alongside corresponding ground-truth and generated masks.
• Figure 3: Continuous iceball region and mask generation for a second cryoablation cycle, starting from an input CT at 0 minutes. Each column depicts the predicted CT slice (top row) and its corresponding iceball mask (bottom row) at incremental time points of +1, +3, +5, and +7 minutes.