Diffusion-driven distillation and contrastive learning for class-incremental semantic segmentation of laparoscopic...

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Accurate identification of anatomical structures in laparoscopic images is vital for surgical procedures.
• Creating specialized datasets for each surgical type is inefficient and challenging.
• Class-incremental semantic segmentation (CISS) offers a potential solution for adapting models to new anatomical classes.

Purpose of the Study:

• To develop a novel algorithm for class-incremental semantic segmentation (CISS) specifically for laparoscopic images.
• To address the limitations of existing CISS methods in clinical settings where incremental data comprises new patient images.
• To improve the adaptability and performance of semantic segmentation models in dynamic surgical environments.

Main Methods:

• A distillation approach driven by a diffusion model was employed for CISS.
• An unconditional diffusion model generated synthetic laparoscopic images for training.
• A distillation network transferred knowledge from previously trained networks.
• Cross-image contrastive learning was utilized to enhance the model's ability to discern subtle anatomical variations.

Main Results:

• The proposed method was evaluated on the Dresden Surgical Anatomy Dataset, encompassing 11 anatomical structures.
• The approach demonstrated superior performance compared to existing methods, particularly for challenging structures like the ureter and vesicular glands.
• Performance surpassed even supervised offline learning in difficult categories.

Conclusions:

• This research presents the first approach to class-incremental semantic segmentation for laparoscopic images.
• The developed method significantly enhances the adaptability of segmentation models to new anatomical classes encountered during surgical procedures.
• The findings pave the way for more robust and versatile AI tools in laparoscopic surgery.