Multi-Scale Hybrid Network for Polyp Detection in Wireless Capsule Endoscopy and Colonoscopy Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Detecting small polyps in wireless capsule endoscopy (WCE) images presents a challenge due to the speed-precision trade-off.
• Medical privacy concerns can limit the acquisition of WCE images, necessitating robust detection methods.

Purpose of the Study:

• To propose a hybrid network, Hyb-SSDNet, for accurate and efficient detection of small polyps in WCE and colonoscopy frames.
• To address limitations in WCE image acquisition by enlarging datasets and employing deep transfer learning.

Main Methods:

• Developed Hyb-SSDNet, a hybrid network combining Inception v4 architecture with a single-shot multibox detector (SSD).
• Incorporated inception blocks to enhance contextual and semantic information capture.
• Utilized multi-scale encoding, weighted feature concatenation, and feature map fusion for improved feature extraction.
• Leveraged deep transfer learning and enlarged datasets for training.

Main Results:

• Achieved a mean average precision (mAP) of 93.29% on the WCE dataset.
• Demonstrated a testing speed of 44.5 frames per second (FPS).

Conclusions:

• Deep learning, specifically the Hyb-SSDNet framework, shows significant potential for advancing polyp detection and classification.
• The proposed method effectively balances speed and precision in identifying small polyps from endoscopic imaging.