To develop a lightweight neural network model for real-time diagnosis and localization in wireless capsule endoscopy (WCE).
Key Findings:
NCA models achieved 29.1% higher accuracy in bleeding segmentation compared to other small-scale models.
NCA produced convincing depth maps from RGB images using pseudo-ground truth data.
The approach allows for real-time segmentation and depth estimation directly on the capsule endoscope.
Interpretation:
The introduction of NCA models represents a significant advancement in the capabilities of capsule endoscopy, enabling more efficient diagnosis and localization without additional hardware.
Limitations:
The scarcity of annotated data for bleeding detection in WCE remains a challenge.
Current models may still require further validation on diverse datasets to ensure robustness.
Conclusion:
eNCApsulate demonstrates the feasibility of integrating advanced neural network models into capsule endoscopes, paving the way for improved diagnostic precision and localization in gastrointestinal imaging.
