Colonic Polyp Detection in Endoscopic Videos With Single Shot Detection Based Deep Convolutional Neural Network

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Colorectal cancer (CRC) prevalence is rising, increasing healthcare costs and mortality.
• Early detection and removal of colonic polyps are crucial for CRC prevention.
• Detecting polyps in colonoscopy videos is challenging due to the complex colon environment and polyp variability.

Purpose of the Study:

• To investigate the effectiveness of the Single Shot Detector (SSD) framework for polyp detection in colonoscopy videos.
• To evaluate different feature extractors (ResNet50, VGG16, InceptionV3) within the SSD framework.
• To compare the performance of the proposed SSD method against existing state-of-the-art methods.

Main Methods:

• The study employed the Single Shot Detector (SSD) framework, a one-stage object detection method using Convolutional Neural Networks (CNNs).
• Three feature extractors (ResNet50, VGG16, InceptionV3) were assessed, with multi-scale feature maps integrated for ResNet50 and InceptionV3.
• The method was validated on the 2015 MICCAI polyp detection challenge datasets and compared with YOLOV3 and Faster-RCNN.

Main Results:

• The proposed SSD method surpassed all competing teams in the MICCAI challenge and YOLOV3 in detection performance.
• The method demonstrated real-time detection speed, outperforming all compared approaches.
• InceptionV3 as a feature extractor yielded the best precision and recall results among the evaluated options.

Conclusions:

• The SSD-based method achieves excellent performance for polyp detection in colonoscopy videos.
• This approach has the potential to significantly improve diagnostic accuracy and efficiency in CRC screening.
• The study highlights the efficacy of SSD with advanced feature extractors for real-time medical image analysis.