边界加权的甲状腺癌病理图像细胞核分割方法

Nuclear segmentation method for thyroid carcinoma pathologic images based on boundary weighting

甲状腺癌是实体癌中发病率增速最快的恶性肿瘤之一,病理学诊断是医生诊断肿瘤的黄金标准,而细胞核分割是病理图像自动分析的关键步骤。针对细胞核分割中细胞核边界位置难以分割问题,设计了边界加权模块使网络在训练时更多关注细胞核边界。另一方面,为了避免网络过分关注边界而忽视细胞核主体部分,导致一些染色较浅的细胞核分割失败,提出了前景增强分割网络;该网络通过在上采样的过程中添加前景增强模块不断增强前景并抑制背景,从而实现细胞核精准分割。在自建的甲状腺癌病理图像分割数据集VIP-TCHis-Seg上的相似系数(Dice)和像素准确率(PA)两个指标分别约为85.26%和95.89%,在公共细胞核分割数据集MoNuSeg上的相似系数(Dice)和像素准确率(PA)两个指标分别约为81.03%和94.63%。上述实验结果表明,提出的边界加权和前景增强模块的方法能有效提高网络在边界处的分割准确率。

Thyroid cancer is one of the most rapidly growing malignancies among all solid cancers.Pathological diagnosis is the gold standard for doctors to diagnose tumors,and nuclear segmentation is a key step in the automatic analysis of pathological images.Aiming at the low segmentation performance of existing segmentation methods on the nuclear boundary of the cell nucleus in the thyroid carcinoma pathological image,we propose an improved U-Net method based on boundary weighting for nuclear segmentation.This method uses the designed boundary weighting module,which can make the segmentation network pay more attention to the boundary of the nuclear.At the same time,in order to avoid the proposed network paying too much attention to the boundary and ignoring the main part of the nucleus,which leads to the failure for some lightly stained nuclei segmentation,we design a segmentation network to enhance the foreground area and suppresses the background area in the upsampling stage.In addition,we build a dataset for nuclear segmentation of thyroid carcinoma pathologic images named VIP-TCHis-Seg dataset.Our method achieves the Dice coefficient(Dice)of 85.26%and the pixel accuracy(PA)of 95.89%on self-built TCHis-Seg dataset,and achieves the Dice coefficient(Dice)of 81.03%and the pixel accuracy(PA)of 94.63%on common dataset MoNuSeg.Experimental results show that our method can achieve the best performance on both Dice and PA as well as effectively improve the segmentation accuracy of the network at the boundary compared with other methods.