Jiang Yun, Dong Jinkun, Zhang Yuan, Cheng Tongtong, Lin Xin, Liang Jing
College of Computer Science and Engineering, Northwest Normal University, Lanzhou, China.
Heliyon. 2023 Feb 26;9(3):e13942. doi: 10.1016/j.heliyon.2023.e13942. eCollection 2023 Mar.
Skin lesion segmentation is a crucial step in the process of skin cancer diagnosis and treatment. The variation in position, shape, size and edges of skin lesion areas poses a challenge for accurate segmentation of skin lesion areas through dermoscopic images. To meet these challenges, in this paper, using UNet as the baseline model, a convolutional neural network based on position and context information fusion attention is proposed, called PCF-Net. A novel two-branch attention mechanism is designed to aggregate Position and Context information, called Position and Context Information Aggregation Attention Module (PCFAM). A global context information complementary module (GCCM) was developed to obtain long-range dependencies. A multi-scale grouped dilated convolution feature extraction module (MSEM) was proposed to capture multi-scale feature information and place it in the bottleneck of UNet. On the ISIC2018 dataset, a large volume of ablation experiments demonstrated the superiority of PCF-Net for dermoscopic image segmentation after adding PCFAM, GCCM and MSEM. Compared with other state-of-the-art methods, the performance of PCF-Net achieves a competitive result in all metrics.
皮肤病变分割是皮肤癌诊断和治疗过程中的关键步骤。皮肤病变区域在位置、形状、大小和边缘上的变化，给通过皮肤镜图像准确分割皮肤病变区域带来了挑战。为应对这些挑战，本文以UNet作为基线模型，提出了一种基于位置和上下文信息融合注意力的卷积神经网络，称为PCF-Net。设计了一种新颖的双分支注意力机制来聚合位置和上下文信息，称为位置和上下文信息聚合注意力模块（PCFAM）。开发了一个全局上下文信息互补模块（GCCM）以获得长程依赖关系。提出了一个多尺度分组扩张卷积特征提取模块（MSEM）来捕获多尺度特征信息并将其置于UNet的瓶颈处。在ISIC2018数据集上，大量的消融实验证明了在添加PCFAM、GCCM和MSEM后，PCF-Net在皮肤镜图像分割方面的优越性。与其他最新方法相比，PCF-Net在所有指标上的性能都取得了具有竞争力的结果。
