摘要
肿瘤细胞的密度、核质比和平均尺寸等指标对癌症的分级和预后有重要的意义.在计算病理学中,细胞核分割是肿瘤微环境分析的基础.通过对分割结果进行统计分析,对新的肿瘤标记物的探索有重大的意义.病理图像背景下的细胞核形态不规则,细胞核染色不均匀,且细胞核边缘之间存在黏连的问题,而现有的深度学习算法在细胞核主体分割正确的情况下,边缘的分割错误不会对总体的损失造成太大的影响,黏连的细胞核很容易被当作同一个分割目标.为了解决细胞核重叠问题,本文提出一种基于Transformer与距离图谱的分割模型,将Transformer中的核心多头自注意力机制与距离图谱引导算法相结合,重视细胞核内部,弱化细胞核边界,提升对图像局部和全局上下文信息的学习能力.本文方法在两个公开数据集上的平均Dice系数为0.7979、精度为0.7561、AJI系数为0.6672、Hausdorff距离为10.11.实验结果表明,相较其他分割算法,本文方法的性能更好,能够有效提高细胞核的分割精度,同时较好地解决了细胞核之间的黏连问题.
Indices such as tumor cell density,nucleocytoplasmic ratio,and average size have important implications for cancer grading and prognosis.Therefore,segmentation of nuclei is the fundamental prerequisite for tumor microenvironment analysis in computational pathology.Additionally,the exploration of new tumor markers is of great significance through statistical analysis of segmentation results.However,the morphology of nuclei in the background of pathological images is irregular,the staining of nuclei is uneven,and adhesion occurs between the edges of nuclei.While the segmentation error of the edge will make no difference on the overall loss as long as the main body of the nucl is correctly segmented,so the adhering nuclei can easily be regarded as the same segmentation target by existing deep learning algorithms.To address the overlapping nuclei,a new segmentation algorithm based on the Transformer and distance map,abbreviated as TDM-Net,is proposed,which integrates the core of multi-head self-attention mechanism in Transformer with contextual information to fully explore the proximity relationship and enhances the learning ability of image details by introducing distance map to emphasize the interior of nuclei and weaken the boundary of nuclei.The algorithm s Dice coefficient,precision,Aggregated Jaccard Index(AJI)and Hausdorff distance are 0.7979,0.7561,0.6672,and 10.11,respectively.The results show that the proposed TDM-Net outperforms other segmentation algorithms,effectively improves nuclei segmentation accuracy and solves overlapping of different nuclei.
作者
鲁浩达
梁实
顾松
王向学
徐军
LU Haoda;LIANG Shi;GU Song;WANG Xiangxue;XU Jun(School of Automation/Institute for Artificial Intelligence in Medicine,Nanjing University of Information Science&Technology,Nanjing 210044,China)
出处
《南京信息工程大学学报（自然科学版）》
CAS
北大核心
2024年第1期66-75,共10页
Journal of Nanjing University of Information Science & Technology（Natural Science Edition）
基金
国家自然科学基金(U1809205,62171230,92159301,62101365,61771249,91959207,81871352)。
