Abstract
Polyp diseases detection in gastrointestinal frames, especially in wireless capsule endoscopy images (WCE) is a challenging task in computer vision and medical imaging purposes. Although many attempt have been done to involve critical human thinking in this field, the precision-vs-speed trade-off still the biggest challenge for recent existing polyp detection networks. Heretofore, WCE images acquisition still provides a challenge owing to the lack of large and publicly available annotated datasets. In order to handle the problem of high missed diagnosis rate of small polyp regions of the gastrointestinal endoscopic images by experts, in this paper, we investigated the potential of single shot multibox detector (SSD) and transfer learning to propose an automated gastrointestinal polyp detection (GP-SSD) method. First, we apply data augmentation strategies. Then, we investigated deep transfer learning transferring knowledge to polyp images. Afterwards, through VGG16 network as a backbone, low-level feature maps are delivered to the next layer, followed by some down-sampling blocks to generate new pyramidal layers. Finally, feature maps are fed to multi-box detectors predicting the final detection results. Other publicly annotated colonoscopy dataset is investigated in order to verify the GP-SSD practicability. Using a proper fine-tuning mechanisms, the GPSSD achieves 79.8% mean average precision (mAP) and 46 frame per second (FPS) on the WCE and colonoscopy datasest. The results indicate that the proposed GP-SSD is more suitable for deployment to embedded devices for real-time object detection. The proposal demonstrates that deep learning has a lot of room for development in the field of gastrointestinal image detection.