Colorectal cancer remains the third most common cancer worldwide, with over 1.9 million diagnosed cases and more than 930,000 deaths in 2020 alone. A critical challenge lies in detecting precancerous colorectal polyps, which vary greatly in size, shape, and appearance. During colonoscopy, even experienced physicians have a miss rate as high as 27% for small polyps.
To address this, a research team led by LI Hailong and LIU Guohua from Donghua University, together with ZHAO Meng from Yanshan University, proposed an improved YOLO-based model named EF-YOLO. The model incorporates several key innovations:

• Advanced multi-scale aggregation (AMSA): replaces the traditional spatial pyramid pooling module to better capture polyps of different sizes.
• Deformable convolutional network-MaxPool (DCN-MP): adaptively samples irregular polyp shapes, preserving critical morphological features.
• Transformer encoder: extracts global contextual information, especially beneficial for small or ambiguous polyps.
• Coordinate attention (CA): enhances focus on polyp regions by integrating positional and channel information.

The model was trained and tested on a merged dataset of 1,612 images from Kvasir-SEG and CVC-ClinicDB. Results show that EF-YOLO achieves a mean average precision (mAP) of 96.60% and a recall of 92.73%, outperforming the baseline YOLOv7 (94.77% mAP, 90.91% recall). In detecting small polyps (areas <5% of the image), the mAP reached 98.86%, demonstrating the model’s exceptional sensitivity.
Moreover, ablation experiments confirmed that each newly introduced module contributed positively to overall performance. The AMSA module alone improved the mAP from 94.77% to 95.94%.
This work highlights the potential of deep learning-based computer-aided diagnosis systems to assist endoscopists in real time, potentially increasing adenoma detection rates and reducing the risk of colorectal cancer progression. The work entitled “An Enhanced Feature Neural Network and Its Application in Detection of Colorectal Polyps” was published in Journal of Donghua University (English Edition) (published in Issue 01, 2026).

DOI: 10.19884/j.1672-5220.202412015