Multi-class motion-based semantic segmentation for ureteroscopy and laser lithotripsy.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Ureteroscopy with laser lithotripsy is a primary kidney stone treatment.
• Accurate segmentation of stones and laser fibers is crucial for automated analysis and surgical guidance.
• Visual challenges like turbidity and motion hinder precise segmentation in ureteroscopy.

Purpose of the Study:

• To develop an automated multi-class segmentation framework for kidney stones and laser fibers in ureteroscopy.
• To enhance surgical decision-making through accurate stone-size estimation and tool identification.
• To overcome visual limitations encountered during laser lithotripsy procedures.

Main Methods:

• Proposed an end-to-end Convolutional Neural Network (CNN) learning framework.
• Utilized a hybrid residual U-Net (HybResUNet) for improved semantic predictions.
• Incorporated a Deformation Vector Field Network (DVFNet) to leverage motion differences for refining segmentation.
• Developed a compound loss function and optimized data augmentation strategies.

Main Results:

• The proposed method achieved superior segmentation performance compared to state-of-the-art models like UNet and DeepLabv3+.
• Demonstrated significant Dice Similarity Coefficient (DSC) improvements on both in vivo and clinical datasets.
• Achieved DSC improvements of 4.15% and 13.34% over UNet and DeepLabv3+ on in vivo data.
• Showcased substantial DSC gains on an unseen clinical dataset, particularly for stone and laser segmentation.

Conclusions:

• The developed CNN framework offers a robust solution for multi-class segmentation in ureteroscopy.
• This technology has the potential to improve surgical outcomes by providing reliable quantitative analysis.
• The study represents a significant advancement in automated analysis for laser lithotripsy procedures.