Colonoscopy polyp detection with massive endoscopic images

TL;DR

This work tackles the challenge of robust, real-time polyp detection in colonoscopy by extending a RetinaNet-based detector with three key enhancements: anchor optimization via differential evolution to improve small-polyp coverage, a degridded dilated-convolution backbone to preserve spatial detail without increasing depth, and an attention gated block to sharpen focus on small, variable polyps. The proposed methods yield measurable gains in detection performance (mAP) on private colonoscopy data, while maintaining real-time inference speed. The study demonstrates that task-specific anchor configurations and light-weight attention mechanisms can meaningfully improve end-to-end polyp localization, suggesting a practical path toward clinically deployable automated screening aids. Overall, the approach reduces clinician workload and enhances screening reliability, with potential for broader clinical validation in real-world settings.

Abstract

We improved an existing end-to-end polyp detection model with better average precision validated by different data sets with trivial cost on detection speed. Our previous work on detecting polyps within colonoscopy provided an efficient end-to-end solution to alleviate doctor's examination overhead. However, our later experiments found this framework is not as robust as before as the condition of polyp capturing varies. In this work, we conducted several studies on data set, identifying main issues that causes low precision rate in the task of polyp detection. We used an optimized anchor generation methods to get better anchor box shape and more boxes are used for detection as we believe this is necessary for small object detection. An alternative backbone is used to compensate the heavy time cost introduced by dense anchor box regression. With use of the attention gate module, our model can achieve state-of-the-art polyp detection performance while still maintain real-time detection speed.

Figures (6)

• Figure 1: Examples of possible polyp category within a frame: (a) sessile, (b) pedunculated or (c) multiples.
• Figure 2: Overall pipeline demonstration for polyp detection
• Figure 3: An illustration of our architecture design: An image will first go to Bottom-up Pathway(Backbone) for feature map accumulation, a Feature Pyramid Network is build upon on the output of Backbone as well as the residual blocks. All stages(P3 to P7) from feature pyramid will be futher shared by two subnets.
• Figure 4: example of anchor optimization. Anchor boxes on the left are the result of optimization, the ratio and sizes offer better coverage with GT label and more boxes are used.
• Figure 5: Layer resolution with dilated convolution.