An artificial intelligence-assisted framework for fast and automatic radiofrequency ablation planning of liver tumors in CT images

To develop and validate an artificial intelligence (AI)-assisted framework for fast and automatic radiofrequency ablation (RFA) planning of liver tumors from CT images.

This framework consisted of three steps. First, the abdominal multi-organs related to RFA planning were automatically segmented from CT images using a modified nnU-Net with class-weighted loss function. Then, utilizing the segmented liver as a location prior, the liver tumors and hepatic vessels were further segmented by a sensitivity-enhanced segmentation network. Finally, a clinically acceptable RFA plan was generated by a fully automatic planning method based on the segmented organs and tumors. Experiments were conducted on two public segmentation datasets and patients from two different centers to evaluate the proposed framework.

The proposed abdominal multi-organ segmentation model achieved an average dice of 87.7 \(\pm\) $\pm$ 8.0% on 15 abdominal organs and the proposed liver tumor and hepatic vessel segmentation model achieved an average dice of 80.7 \(\pm\) $\pm$ 11.2% and 65.3 \(\pm\) $\pm$ 10.2% and an average sensitivity of 87.4 \(\pm\) $\pm$ 12.9% and 76.8 \(\pm\) $\pm$ 14.9% for liver tumor and hepatic vessel, respectively. Finally, the proposed framework generated clinically acceptable RFA plans within a few minutes without human intervention for all patients from two centers.

The proposed AI-assisted framework is fast and can automatically generate clinically acceptable RFA plans for liver tumors from CT images, which can assist interventional radiologists in determining suitable plans and reduce their burden.