Negative Samples for Improving Object Detection-A Case Study in AI-Assisted Colonoscopy for Polyp Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Deep learning object-detection models are crucial for computer-aided diagnosis in colonoscopy polyp detection.
• Existing models often lack negative samples containing common artifacts, leading to reduced real-world performance.

Purpose of the Study:

• To demonstrate the necessity of incorporating negative samples with artifacts into training datasets for polyp detection models.
• To improve the accuracy and reliability of deep learning-based polyp detection systems.

Main Methods:

• Retrained a YOLOv3-based object-detection model using a dataset augmented with 15% non-polyp images featuring various artifacts.
• Evaluated model performance using internal test datasets and four public colonoscopy datasets.

Main Results:

• The retrained model showed improved F1 performance on internal test datasets (0.869 to 0.893).
• Performance also increased on public datasets containing non-polyp images (average F1 from 0.695 to 0.722).
• Inclusion of artifact-rich negative samples reduced false positives and provided more realistic performance estimates.

Conclusions:

• Augmenting training data with diverse negative samples is essential for enhancing the robustness of deep learning polyp detection models.
• This strategy leads to more accurate computer-aided diagnosis systems for colonoscopies, improving patient outcomes.