U-NetPlus: A Modified Encoder-Decoder U-Net Architecture for Semantic and Instance Segmentation of Surgical...

Area of Science:

• Medical technology
• Computer vision
• Robotics

Background:

• Robot-assisted surgery represents a significant advancement in minimally invasive procedures.
• Accurate tracking and identification of surgical instruments are critical challenges in robotic surgery.
• Deep learning semantic segmentation offers a promising approach for analyzing surgical video frames.

Purpose of the Study:

• To develop an improved deep learning model for semantic segmentation of surgical instruments in robotic surgery videos.
• To enhance the detection and identification capabilities of surgical tools within complex surgical scenes.
• To achieve superior performance compared to existing methods on benchmark datasets.

Main Methods:

• Modification of the U-Net architecture, incorporating a pre-trained encoder.
• Redesign of the decoder using nearest-neighbor (NN) interpolation for upsampling.
• Implementation of a fast and flexible data augmentation technique.
• Training and testing the U-NetPlus framework on the MICCAI 2017 EndoVis Challenge dataset.

Main Results:

• Achieved 90.20% DICE score for binary segmentation of surgical instruments.
• Obtained 76.26% DICE score for instrument part segmentation.
• Reached 46.07% DICE score for instrument type segmentation.
• Demonstrated superior performance over previous techniques on the same dataset.

Conclusions:

• The proposed U-NetPlus architecture significantly improves semantic segmentation of surgical instruments in robotic surgery videos.
• The modifications to the U-Net model enhance accuracy in detecting and classifying surgical tools.
• This approach holds potential for advancing computer-assisted surgery and improving surgical outcomes.