Xu Chao, Zhang Yi, Fan Xianjun, Lan Xingjie, Ye Xin, Wu Tongning
China Telecommunication Technology Labs, China Academy of Information and Communications Technology, Beijing, China.
Department of Product Development, Zhuhai Sanmed Biotech Ltd., Zhuhai, China.
Quant Imaging Med Surg. 2022 May;12(5):2961-2976. doi: 10.21037/qims-21-909.
Circulating tumor cells (CTCs) acting as "liquid biopsy" of cancer are cells that have been shed from the primary tumor, which cause the development of a secondary tumor in a distant organ site, leading to cancer metastasis. Recent research suggests that CTCs with abnormalities in gene copy numbers in mononuclear cell-enriched peripheral blood samples, namely circulating genetically abnormal cells (CACs), could be used as a non-invasive decision tool to detect patients with benign pulmonary nodules. Such cells are identified by counting the fluorescence signals of fluorescence in situ hybridization (FISH). However, owing to the rarity of CACs in the blood, identification of CACs using this technique is time-consuming and is a drawback of this method.
This study has proposed an efficient and automatic FISH-based CACs identification approach which is based on a combination of the high accuracy of You Only Look Once (YOLO)-V4 and the lightweight and rapidness of MobileNet-V3. The backbone of YOLO-V4 was replaced with MobileNet-V3 to improve the detection efficiency and prevent overfitting, and the architecture of YOLO-V4 was optimized by utilizing a new feature map with a larger scale to enable the enhanced detection ability for small targets.
We trained and tested the proposed model using a dataset containing more than 7,000 cells based on five-fold cross-validation. All the images in the dataset were 2,448×2,048 (pixels) in size. The number of cells in each image was >70. The accuracy of four-color fluorescence signals detection for our proposed model were all approximately 98%, and the mean average precision (mAP) were close to 100%. The final outcome of the developed method was the type of cells, i.e., normal cells, CACs, gaining cells or deletion cells. The method had a CACs identification accuracy of 93.86% (similar to an expert pathologist), and a detection speed that was about 500 times greater than that of a pathologist.
The developed method could greatly increase the review accuracy, enhance the efficiency of reviewers, and reduce the review turnaround time during CACs identification.
循环肿瘤细胞（CTC）作为癌症的“液体活检”，是从原发性肿瘤脱落的细胞，可导致远处器官部位出现继发性肿瘤，从而引发癌症转移。最近的研究表明，在单核细胞富集的外周血样本中基因拷贝数异常的循环肿瘤细胞，即循环基因异常细胞（CAC），可作为一种非侵入性决策工具来检测良性肺结节患者。此类细胞通过计数荧光原位杂交（FISH）的荧光信号来识别。然而，由于血液中CAC数量稀少，使用该技术识别CAC耗时，这是该方法的一个缺点。
本研究提出了一种基于FISH的高效自动CAC识别方法，该方法结合了You Only Look Once（YOLO）-V4的高精度以及MobileNet-V3的轻量级和快速性。用MobileNet-V3替换YOLO-V4的主干以提高检测效率并防止过拟合，并利用更大尺度的新特征图优化YOLO-V4的架构，以增强对小目标的检测能力。
我们使用一个包含7000多个细胞的数据集，基于五折交叉验证对所提出的模型进行训练和测试。数据集中所有图像的大小均为2448×2048（像素）。每张图像中的细胞数量>70。我们所提出模型对四色荧光信号的检测准确率均约为98%，平均精度均值（mAP）接近100%。所开发方法的最终结果是细胞类型，即正常细胞、CAC、增益细胞或缺失细胞。该方法对CAC的识别准确率为93.86%（与专家病理学家相近），检测速度比病理学家快约500倍。
所开发的方法可大幅提高审查准确率，提高审查人员的效率，并减少CAC识别过程中的审查周转时间。
