Improving Automatic Polyp Detection Using CNN by Exploiting Temporal Dependency in Colonoscopy Video.

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Automatic polyp detection in colonoscopy is challenging due to diverse polyp appearances and similar non-polyp structures.
• Current convolutional neural networks (CNNs) struggle with noise, missed detections in adjacent frames, and high false positive rates.

Purpose of the Study:

• To enhance the performance of CNN-based object detectors for polyp detection in colonoscopy videos.
• To reduce false positives and improve overall detection accuracy, sensitivity, precision, and specificity.

Main Methods:

• A two-stage approach was developed, beginning with region of interest (RoI) proposal using CNNs.
• A false positive (FP) reduction unit was implemented, leveraging bidirectional temporal information from consecutive video frames.

Main Results:

• The integration of bidirectional temporal information significantly improved polyp position estimation and accurate FP prediction.
• The proposed method demonstrated superior performance in sensitivity, precision, and specificity compared to conventional methods.
• State-of-the-art results were achieved on the CVC-ClinicVideoDB video dataset.

Conclusions:

• Exploiting temporal dependencies in colonoscopy videos is effective for reducing false positives in polyp detection.
• The developed two-stage method offers a robust solution for improving automated polyp detection accuracy.