3DGPS: A 3D Differentiable-Gaussian-Based Planning Strategy for Liver Tumor Cryoablation

Ce Wang, Xiaoyu Huang, Yaqing Kong, Qian Li, You Hao, Xiang Zhou

Abstract: Effective preoperative planning is crucial for successful cryoablation of liver tumors. However, conventional planning methods rely heavily on clinicians’ experience, which may not always lead to an optimal solution due to the intricate 3D anatomical structures and clinical constraints. Lots of planning methods have been proposed, but lack interactivity between multiple probes and are difficult to adapt to diverse clinical scenarios. To bridge the gap, we present a novel 3D Differentiable-Gaussian-based Planning Strategy (3DGPS) for cryoablation of liver tumor considering both the probe interactivity and several clinical constraints. Especially, the problem is formulated to search the minimal circumscribed tumor ablation region, which is generated by multiple 3D ellipsoids, each from one cryoprobe. These ellipsoids are parameterized by the differentiable Gaussians and optimized mainly within two stages, fitting and circumscribing, with formulated clinical constraints in an end-to-end manner. Quantitative and qualitative experiments on LiTS and in-house datasets verify the effectiveness of 3DGPS.