Enhanced segmentation of gastrointestinal polyps from capsule endoscopy images with artifacts using ensemble learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Capsule endoscopy (CE) images frequently contain artifacts, unlike standard high-quality datasets.
• These artifacts negatively impact the performance of polyp segmentation and detection models.
• Existing models struggle with the real-world variability of CE imaging.

Purpose of the Study:

• To enhance polyp segmentation accuracy in capsule endoscopy (CE) images affected by artifacts.
• To investigate the efficacy of ensemble learning for improving polyp detection in challenging CE images.
• To compare the performance of ensemble models against single models on real-world CE data.

Main Methods:

• Collected 277 polyp images with CE artifacts from 480 patients.
• Utilized two public high-quality external datasets for comparative analysis.
• Trained and evaluated base models and an ensemble model for polyp segmentation.

Main Results:

• Artifacts in CE images degraded the performance of single semantic segmentation models.
• The ensemble model outperformed the best single models on real-world CE datasets with artifacts.
• Ensemble learning demonstrated notable increases in Intersection over Union (IoU) and Dice scores (0.08%-7.01% and 0.61%-4.93%) on artifact-containing data.

Conclusions:

• Ensemble learning is an effective strategy for improving polyp segmentation accuracy in artifact-laden CE images.
• The proposed method enhances polyp detection rates, even in the presence of significant imaging artifacts.
• This approach offers a robust solution for analyzing real-world capsule endoscopy data.