Self-Supervised Lightweight Depth Estimation in Endoscopy Combining CNN and Transformer.

Area of Science:

• Medical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• 3D reconstruction is crucial for medical engineering tasks like surgical navigation and robotics.
• Self-supervised depth estimation is valuable for endoscopic procedures, avoiding the need for ground truth data.
• Existing methods often require large parameter counts, motivating the development of efficient models.

Purpose of the Study:

• To propose a lightweight, self-supervised depth estimation network for medical imaging.
• To enhance feature extraction by tightly coupling Convolutional Neural Networks (CNNs) and Transformers.
• To improve pose prediction accuracy through multi-head attention mechanisms.

Main Methods:

• A novel network architecture integrating CNN and Transformer modules at different encoder scales.
• Utilizing CNNs for local texture perception and Transformers for global shape extraction within a hierarchical structure.
• Incorporating multi-head attention modules into the pose network for enhanced accuracy.

Main Results:

• The proposed lightweight network achieves comparable performance to existing methods on two datasets.
• The model effectively compresses parameters, offering a more efficient solution for depth estimation.
• The hierarchical feature extraction leverages the complementary strengths of CNNs and Transformers.

Conclusions:

• The developed lightweight CNN-Transformer network provides an effective solution for self-supervised depth estimation in medical applications.
• This approach demonstrates the potential of tightly coupled CNN-Transformer architectures for efficient and accurate 3D reconstruction.
• The method offers a promising direction for advancing medical imaging technologies requiring precise depth information.