Self-supervised monocular depth estimation for gastrointestinal endoscopy.

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Gastrointestinal (GI) endoscopy is vital for cancer screening but faces challenges in polyp detection and lesion follow-up due to limited field of view and varying endoscopist skill.
• Accurate depth estimation from GI endoscopic videos is crucial for developing AI-assisted surgical techniques.
• Existing depth estimation algorithms struggle with the unique environment and data limitations of GI endoscopy.

Purpose of the Study:

• To propose a self-supervised monocular depth estimation method specifically designed for GI endoscopic sequences.
• To improve the accuracy of depth estimation in GI endoscopy for better lesion identification and surgical assistance.

Main Methods:

• Developed a depth estimation network and a camera ego-motion estimation network for self-supervised training.
• Utilized a multi-scale structural similarity with L1 norm (MS-SSIM+L1) loss function to preserve high-frequency details and color invariance.
• Employed a U-shape convolutional network with a dual-attention mechanism to capture multi-scale contextual information.

Main Results:

• The proposed method demonstrated superior generality, achieving lower error and higher accuracy metrics on both the UCL and Endoslam datasets.
• Qualitative and quantitative evaluations showed competitive performance against state-of-the-art methods.
• The model's effectiveness was validated in clinical GI endoscopy settings.

Conclusions:

• The developed self-supervised depth estimation method shows significant potential for clinical application in GI endoscopy.
• This technique can enhance AI-assisted tools for more accurate polyp detection and precancerous lesion management.
• The method offers a promising advancement for improving GI cancer screening and endoscopic procedures.