面向头颈部肿瘤转移性淋巴结分割网络

Segmentation network for metastatic lymph nodes of head and neck tumors

头颈部肿瘤是我国常见的恶性肿瘤,其预后主要受颈部淋巴结转移的影响,医学上通过核磁共振成像技术对转移性淋巴结成像后再进行诊断,然而,核磁共振技术成像存在病灶形态信息丢失,病灶区域对比度低和病灶边界模糊的问题。针对这些问题,提出面向头颈部肿瘤转移性淋巴结分割网络协助医生进行诊断。首先,设计跨层跨视野注意力模块,其接收深浅层的特征信息后利用自注意力机制分别突显深浅层的转移淋巴形状,通过不同感受野的深层特征图学习到更好的语义上下文特征,将浅层特征图与深层特征图逐像素融合,增强转移淋巴病灶区域的形态信息。其次,设计多尺度特征融合模块,在特征金字塔的初始位置融合不同尺度的特征图,丰富转移淋巴病灶区域的形态信息。然后,设计增强注意力预测头模块,通过对预测前的特征图使用并行的自注意力与门控通道转换模块,凸出病灶区域,细化病灶边界。最后,使用临床淋巴结转移医学图像数据集验证网络的有效性。实验结果表明,所提网络对于淋巴结转移病灶分割的APdet,APseg,ARdet,ARseg,mAPdet和mAPseg分别为74.88%,74.12%,63.11%,62.28%,74.64%和74.04%。该网络实现对淋巴结转移病灶区域的精确检测分割,对辅助淋巴结诊断具有积极意义。

Head and neck tumors are prevalent malignant tumors in China,with prognosis significantly in⁃fluenced by cervical lymph node metastasis.In medical practice,magnetic resonance imaging(MRI)is em⁃ployed to identify metastatic lymph nodes.However,MRI images often suffer from blurred edges and low contrast between the lesion and surrounding tissue. This paper introduces a segmentation network tailoredfor metastatic lymph nodes in head and neck tumors. Initially, a cross-layer and cross-field attention mod⁃ule is developed, integrating features from both deep and shallow layers to enhance the shape representa⁃tion of metastatic lymph nodes through a self-attention mechanism. This module improves contextual se⁃mantic understanding across different receptive fields, allowing for pixel-level fusion of shallow and deepfeature maps, thereby enhancing the morphological details of metastatic lymphatic nodes. Subsequently, amulti-scale feature fusion module is designed to amalgamate features across various scales in the featurepyramid, enriching the morphological details of the lymph nodes. Furthermore, an enhanced attention pre⁃diction head module is implemented, combining parallel self-attention and gate channel transformation toaccentuate the lesion area and refine its boundaries on the feature map. The network's effectiveness is con⁃firmed using a clinical dataset of lymph node metastasis medical images. The performance metrics, AP⁃det, APseg, ARdet, ARseg, mAPdet, and mAPseg for lymph node metastasis lesion segmentation are74.88%, 74.12%, 63.11%, 62.28%, 74.64%, and 74.04%, respectively. This network provides pre⁃cise detection and segmentation of lymph node metastasis lesions, offering significant benefits for lymphnode diagnosis.