SurgAI: deep learning for computerized laparoscopic image understanding in gynaecology

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence in Surgery

Background:

• Laparoscopic surgery can benefit from image-guided systems for enhanced visualization.
• Current artificial intelligence (AI) in gynecology lacks precise anatomical localization, offering only presence detection.
• Semantic segmentation offers pixel-level localization, crucial for advanced image understanding in surgery.

Purpose of the Study:

• To introduce the first dedicated dataset for deep learning-based semantic segmentation in gynecological laparoscopy.
• To evaluate the performance of deep learning models for semantic segmentation of anatomical structures and surgical tools in laparoscopic images.
• To advance image understanding for potential integration into image-guided surgery systems.

Main Methods:

• Utilized the Mask R-CNN deep learning model.
• Developed a dataset of 461 manually annotated laparoscopic images (uterus, ovaries, surgical tools).
• Trained Mask R-CNN on 361 images and evaluated on 100 images.

Main Results:

• Achieved segmentation accuracies of 84.5% (uterus), 29.6% (ovaries), and 54.5% (surgical tools).
• Inferred detection accuracies of 97% (uterus), 24% (ovaries), and 86% (surgical tools).
• Demonstrated state-of-the-art detection performance for uterus and surgical tools, with lower performance for ovaries.

Conclusions:

• Preliminary results are promising, despite the initial dataset size.
• Deep learning semantic segmentation shows potential for improving gynecological laparoscopic surgery.
• The development of an international surgical image database is recommended for future advancements.