基于深度学习框架的循环染色体异常细胞识别  

Identification of circulating genetically abnormal cell based on deep learning framework

作　　者：徐旭 李从胜 王祺轩 范献军 巫彤宁 XU Xu;LI Congsheng;WANG Qixuan;FAN Xianjun;WU Tongning(China Telecommunication Technology Labs,China Academy of Information and Communications Technology,Beijing 100191;Zhuhai Sanmed BIOTECH Ltd.,Zhuhai,Guangdong Province 519060)

机构地区：[1]中国信息通信研究院泰尔终端实验室,北京100191 [2]珠海圣美生物诊断技术有限公司,广东珠海519060

出　　处：《北京生物医学工程》2023年第6期551-558,共8页Beijing Biomedical Engineering

基　　金：国家自然科学基金(61971445、62271508)资助。

摘　　要：目的 循环染色体异常细胞(circulating genetically abnormal cell,CAC)检测是一种无创或微创、敏感、经济的肿瘤早期诊断方法。荧光原位杂交(fluorescence in situ hybridization,FISH)通过计算荧光探针在细胞核内产生的增益,可以准确地检测CAC中基因异常的繁殖。然而,基于FISH的CAC识别存在细胞核重叠和荧光信号形态多样的问题,人工检测荧光信号是困难的。方法 本研究基于四色荧光原位杂交肺癌图像,针对细胞核和染色信号形态特征提出一种级联细胞核分割网络(cascaded nuclei segmentation network,CACNET)和信号点检测网络(signal detection network,FISHNET)的CAC识别方案。首先使用CACNET对细胞核进行分割获取掩膜,并对应裁剪荧光染色信号图像,然后对单细胞核内的信号点使用FISHNET进行检测,最终对信号点计数并利用临床诊断方法判别细胞。其中,CACNET通过结合注意力机制和边缘约束算法提升重叠细胞核分割效果。FISHNET通过热图回归拟合染色信号形态,并且自定义轻量化目标检测网络提升信号点识别性能。结果 本研究构建44 000细胞核进行模型的训练和测试。测试结果显示,本研究提出的CACNET在细胞核分割上取得92.67%的交并比,FISHNET在信号点检测的平均精度接近98%且模型参数量仅有2.24 M。CAC识别的敏感性和特异性分别为96.52%和99.1%。结论 本研究提出的CAC识别方案是高效且鲁棒的,并且与现有的CAC识别方法相比有一定优势,有助于提高临床肺癌诊断结果的可靠性。Objective The detection of circulating genetically abnormal cell(CAC) is a non-invasive or minimally invasive,sensitive,and cost-effective method for the early diagnosis of tumors.Fluorescence in situ hybridization(FISH) is able to accurately detect the propagation of genetic abnormalities in CAC by calculating the gain generated by the fluorescent probe in the nucleus.However,FISH-based CAC identification suffers from the problems of overlapping cell nuclei and diverse fluorescent signal morphology,and manual detection of fluorescent signals is difficult.Methods We propose a CAC identification scheme based on 4-color fluorescence in situ hybridization lung cancer images with cascaded nuclei segmentation network(CACNET) and signal detection network(FISHNET) for the morphological features of nuclei and staining signals.Firstly,the nuclei are segmented by using CACNET to obtain masks and correspondingly cropped fluorescent stained signal images.Then the signal spots within single cell nuclei are detected by using FISHNET,and finally the signal spots are counted and the cells are discriminated by using clinical diagnostic methods.Among them,CACNET enhances the overlapping cell nuclei segmentation through combining attention mechanism and edge constraint algorithm.FISHNET improves signal point identification performance by fitting stained signal morphology through heat map regression and customizing a lightweight target detection network.Results The CACNET achieve 92.67% on intersection over union(IoU) in nucleus segmentation,and the average precision(AP) of FISHNET in signal point detection is up to 98%,with only 2.24 M model parameters.The sensitivity and specificity of CAC identification are 96.52% and 99.1%,respectively.Conclusions The CAC identification scheme proposed in this study is efficient and robust,and has advantages over existing CAC identification methods,which could help to improve the reliability of clinical lung cancer diagnosis results.

关 键 词：液体活检 细胞识别 深度学习 细胞核分割 目标检测 

分 类 号：R318.04[医药卫生—生物医学工程]