Streamlining Renal Tumor Ablation Planning with a Deep-Learning Pipeline for Automated Tumor Segmentation and Ablation Zone Optimization

Purpose: Kidney cancer represents about 4% of U.S. cancer diagnoses and 2% of global cases and deaths. Image-guided thermal ablation (IGTA) is a minimally invasive treatment for small renal masses, offering advantages over surgery and achieving comparable long-term outcomes.  However, image processing techniques involved in IGTA pre-procedural planning can be time-consuming and variable, underscoring a need for automation and standardization. This study aims to develop a deep learning (DL) algorithm for automatic kidney tumor segmentation and to optimize the selection of ablation probe parameters to improve precision, accuracy, and standardization in tumor ablation planning.

Materials and Methods: 489 computed tomography (CT) scans from the KiTS23 public dataset and 78 scans from Rhode Island Hospital (RIH) were used for model training and testing. A coarse-to-fine segmentation approach was implemented using a 3D ResUNet model for coarse kidney segmentation, followed by fine tumor segmentation using the MONAI Auto3DSeg framework with 5-fold cross-validation. For each tumor segmentation, Pareto optimization was applied to select optimal ablation zones from 2115 candidates on the basis of ablation zone volume and minimum, mean, and Q1 treatment margins. Margin estimates achieved by manual segmentation and probe selection by human experts were compared to those generated by the Pareto optimization algorithm using paired t-tests.

Results: The DL algorithm achieved a tumor Dice coefficient of 0.7352 on a held-out internal test set (n = 48) and 0.5515 on an external test set (n = 78). The optimization algorithm identified Pareto efficient combinations of probe, power, time, and orientation parameters. In a cohort of 36 RIH patients who underwent IGTA, the algorithm-generated ablation zone plans achieved an average minimum treatment margin of 5.708 mm (SD = 0.934) compared to 3.410 mm (SD = 1.486) margin by a human annotator (t = -35.356, p < 0.001). For volume-normalized minimum treatment margins, the algorithm averaged 0.184 (SD = 0.0290), while the human expert averaged 0.131 (SD = 0.0575) (t = -27.536, p < 0.001).

Conclusion: This project developed and evaluated an end-to-end DL pipeline for improving the precision and throughput of IGTA planning. This study demonstrates the potential of such a pipeline to enhance the efficiency and reproducibility of the IGTA planning process while achieving estimated treatment margins comparable to or better than human experts. However, limitations of this work include generalizability to external datasets and consideration of anatomy beyond the kidney.