Generative AI-Assisted Indication of Anatomical Landmarks to Enhance Safety under Intraoperative Bleeding in...

This study aimed to present a novel approach that integrates a generative adversarial network (GAN) into the inference process of surgical navigation artificial intelligence (AI) systems.

Safe laparoscopic gastrectomy (LG) for gastric cancer relies on accurate identification of anatomical landmarks, such as the pancreas and dimpling lines (DLs). We developed a gastric cancer/gastrectomy AI (gAI) to indicate these landmarks. However, intraoperative bleeding often obscures the operative field, reducing the effectiveness of gAI.

We created gAI-with-GAN by integrating a CycleGAN-based generative model into gAI to remove intraoperative bleeding during the inference process. We compared gAI-with-GAN with gAI using 30 LG videos categorized by bleeding severity. Expert surgeons evaluated the software usefulness, using a 5-point Likert scale and Dice coefficients.

In massive bleeding cases, gAI-with-GAN outperformed gAI. The median Likert scores for identifying the pancreas and the DL between the mesogastrium and pancreas improved from 3 to 4 (P < 0.05). Scores for the accuracy of pancreatic indication improved from 2 to 3 (P < 0.05), and scores for potential clinical utility in preventing postoperative pancreatic fistula increased from 2 to 4 (P < 0.005). Dice coefficients for the pancreas increased from 0.36 ± 0.26 to 0.65 ± 0.17 (P < 0.05).

Integrating generative AI improved the quality of anatomical landmark indications, even under intraoperative bleeding during LG. This approach may enhance surgical safety in minimally invasive surgeries.