Details preserved unsupervised depth estimation by fusing traditional stereo knowledge from laparoscopic images

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Navigation

Background:

• Depth estimation is crucial for vision-based laparoscope surgical navigation.
• Obtaining ground truth depth data for training is challenging in laparoscopic procedures.

Purpose of the Study:

• To develop an unsupervised learning approach for accurate depth estimation in laparoscopy.
• To fuse traditional stereo knowledge with deep learning to address data limitations.

Main Methods:

• Generated proxy disparity labels using a traditional stereo method with confidence measures.
• Trained a dual encoder-decoder convolutional neural network on rectified stereo images and proxy labels.
• Implemented a principled mask to exclude parallax-affected pixels and a neighborhood smoothness term for surface consistency.

Main Results:

• The unsupervised method achieved accurate depth estimation without ground truth data.
• The approach preserved realistic surgical site details and improved point cloud accuracy.
• Validated performance on da Vinci partial nephrectomy and Hamlyn Centre heart phantom datasets.

Conclusions:

• The proposed unsupervised depth estimation method effectively addresses data acquisition challenges in laparoscopy.
• This technique enhances the accuracy and realism of depth maps for surgical navigation.
• The fusion of traditional stereo and deep learning offers a promising direction for medical imaging applications.