基于频率特征增强的结直肠息肉分割模型

Colon Polyp Segmentation Model Based on Frequency Feature Enhancement

摘要: 早期息肉检查是防范结直肠癌发病的重要手段，针对现有基于深度学习方法依旧不能准确辨别息肉位置和边缘信息的问题，提出了一种利用傅里叶变换增强频率特征(FFENet)的息肉分割方法。具体地，在FFENet中设计了一个细节特征增强注意力模块和一个全局频率特征学习模块，前者重耦合不同深度的特征并计算三种显著性特征图来细化息肉区域及其边缘；后者在频域中引入可学习的滤波核，以增强息肉与其边缘间的连贯性并捕捉图像像素之间的长距离依赖关系。结合改善的部分解码器和自适应特征选择模块，大量实验结果表明所提出FFENet在五类息肉数据上更具优势。尤其是在ETIS数据集上，对比其他最先进的模型，大模型版本FFENet-L在Dice和IoU指标上分别提升了4%和5.5%，而小模型版本FFENet-S在保持精度相当的同时，仅仅使用了6.2M参数。

Abstract: Colorectal cancer mainly originates from mutated polyp tissues, so early polyp examination can effectively reduce the incidence of colorectal cancer. The ability to automatically and accurately assist doctors in screening polyps is of great significance in the clinical diagnosis of colorectal cancer. However, existing deep learning-based methods cannot fully address the challenges posed by the color, brightness, and complexity of polyps in polyp images, making it difficult to accurately distinguish the location and edge information of polyps. Considering these problems, we innovatively utilize the Fourier Transform in traditional digital image processing to propose a frequency feature enhancement network (FFENet) for polyp segmentation. In FFENet, a detail feature enhancement attention module and a global frequency feature learning module are presented based on the frequency features. DFEA recouples features from different depths and calculates three types of saliency feature maps to refine the polyp region and polyp boundary. GFFLM aims to enhance the coherence of the polyp body and boundary by incorporating a learnable filtering kernel in the frequency domain to capture the long-range relationships between image pixels. Combined with an enhanced partial decoder and an adaptive feature selection module, our method excels in small polyp segmentation and polyp segmentation in complex environments. Extensive experiments are conducted on five public datasets and demonstrate the superiority of our proposed method for polyp segmentation compared with several state-of-the-art methods. Especially on the ETIS dataset, compared to other SOTA models, our large model version FFENet-L has improved by 4% and 5.5% in terms of Dice and IoU metrics respectively, while our small model version FFENet-S has only used 6.2M parameters while maintaining similar accuracy.