CE-YOLO: A channel-efficient YOLO algorithm for colorectal polyp detection

Abstract

Colorectal cancer (CRC) represents a substantial public health challenge, with early detection of precancerous lesions, notably polyps, via colonoscopy being pivotal for timely intervention. Many deep learning-based polyp detection methods have been proposed to improve the polyp detection rate, but these methods are often too complex for clinical edge deployment. In this study, we propose a lightweight, channel-efficient YOLO algorithm, termed CE-YOLO, specifically designed for detecting colorectal polyps. We present the PCST module for feature extraction, develop the DCBM structure for feature fusion and transmission, and create an appropriate detection head. The PCST module combines spatial transformation convolution with a cross-stage partial connection structure and incorporates partial channel convolution. This integration allows for better capture and retention of fine-grained features while significantly reducing computational complexity without compromising detection performance. The DCBM is a multi-branch, multi-scale feature fusion structure utilizing dynamic convolution kernels, facilitating efficient fusion and propagation of multi-scale features along the channel dimension. Additionally, we propose and apply a mixed-precision quantization strategy based on sensitivity traversal analysis, a first in polyp detection, reducing model size for efficient edge deployment while maintaining accuracy. We conduct experiments on our proposed method and existing state-of-the-art object detection algorithms using six datasets, to better evaluate detection performance and generalization capability. The experimental results demonstrate that our model achieves superior performance with the lowest complexity, validating the efficacy and benefits of our approach. Compared to existing studies, our research emphasizes efficiency and lightweight, offering greater potential for clinical application.