**6.2 Artificial intelligence-driven segmentation**

To plan multiple trajectories, especially in a percutaneous fashion, multiple organs surrounding the pancreas need to be considered. To obtain a good spatial understanding of the acquired tomographic images, 3D representations from the structures of interest help in the decision-making process. Where time may be sufficiently available in the preoperative phase, it is a valuable asset during the intraoperative phase. Therefore, a compromise between accuracy and processing time must be chosen to enable 3D reconstruction from intraoperative tomographic images. Deep learning-based segmentation has shown its potential in multiple fields including medical imaging tasks [49]. Segmentation of the pancreas is seen as a challenging task due to the dependency on good delineation of the organ to its adjacent structures in the abdominal cavity [50]. To segment the complete range of structure needed for pancreas IRE planning, Roth et al. [51] describe a multi-scale pyramid of 3D fully convolutional network. The reported network was trained and validated using 377 clinical CT datasets with annotated organs and vessels. The performance of the system was measured by means of the DICE score, which represents the similarity between the automatic segmentations from CT images not used for training and their manual annotations. The network was able to achieve a Dice score close to 90% on average, which holds promising aspects for automatic, intraoperative segmentation [51].

### **6.3 Automatic trajectory optimization**

Based on the patient-specific data, automatic trajectories can be computed according to permitted access paths while securing a specific distance to structures at risk [52]. The optimization for the IRE use case would further depend on the needle configuration, interelectrode spacing, and parallelism constraints.
