首页|基于混合编码的皮肤病变图像分割

基于混合编码的皮肤病变图像分割

扫码查看
原文链接
  • 万方数据
  • 维普
皮肤镜图像中的皮肤病变分割是计算机辅助诊断皮肤癌的关键.现有的卷积神经网络(CNNs)通常由于卷积操作的固有限制而忽视全局上下文信息.因此,提出了一种具有Transformer和CNN优点的混合编码器的皮损分割网络.首先,使用极坐标变换对原始图像进行预处理.接着利用CNN对不同尺度特征进行预提取,并将其作为Transformer编码器的输入序列,实现对序列数据的全局上下文建模,更好地捕获特征之间的长程依赖关系.最后,在解码器中加入了多级特征融合模块和注意力机制,解码不同尺度和编码块内的分层语义特征.提出的HET-Net网络在ISIC 2018数据集上JSI、DSC和ACC值分别达到了85.09％、91.43％和96.90％,在ISIC 2016+PH2数据集上分别达到了87.44％、93.02％和95.68％.与其他模型相比,所提模型取得了显著的结果,验证了模型的有效性.
HET-Net:Skin Lesion Image Segmentation Based on Hybrid Encoding
Skin lesion segmentation from dermoscopy images is crucial for computer-aided diagnosis of skin cancer. Existing convolution-al neural networks( CNNs) often overlook global contextual information due to the inherent limitations of convolutional operations. There-fore,a hybrid encoder for skin lesion segmentation is proposed that combines the advantages of both a transformer and CNN. To prepro-cess the original images,a polar coordinate transformation is applied. Next,CNN is used to extract features at different scales and learn local image characteristics. The feature sequence extracted by CNN serves as the input to the transformer encoder,enabling global con-text modeling of the sequence data and better capturing long-range dependencies between features. In the decoder,a multi-level feature fusion module and attention mechanism are incorporated to decode features of different scales and hierarchical semantics within the en-coding blocks. The proposed HET-Net network achieves JSI,DSC,and ACC values of 85.09％,91.43％,and 96.90％ respectively on the ISIC 2018 dataset,and 87.44％,93.02％,and 95.68％ respectively on the ISIC 2016+PH2 dataset. Compared to other models,the pro-posed model demonstrates significant results,validating its effectiveness.

skin lesion segmentationTransformerhybrid encodermultilevel feature cascade fusion

彭静、马玉良、席旭刚

展开 >

杭州电子科技大学自动化学院,浙江 杭州310018

皮损分割 Transformer 混合编码 多级特征级联融合

国家科技部科技创新重大项目(2030)国家自然科学基金国家自然科学基金浙江省教育厅科研项目浙江省教育厅科研项目浙江省教育厅科研项目杭州电子科技大学研究生科研创新基金

2021ZD01132046207116161372023Y202250095Y202351775Y202351785CXJJ2022149

2024

10.3969/j.issn.1004-1699.2024.03.015

传感技术学报
东南大学 中国微米纳米技术学会

传感技术学报

CSTPCD北大核心
影响因子：1.276
ISSN：1004-1699
年,卷(期)：2024.37(3)
  • 2