目的:口腔癌是人类第六常见的癌症。用于计数嗜卵核仁组织区域(AgNORs)的刷细胞学可以帮助早期发现口腔癌,降低患者死亡率。然而,今天仍在使用的AgNOR手动计数是耗时的,劳动密集型,而且容易出错。我们工作的目标是通过提出一种基于卷积神经网络(CNN)的方法来解决这些缺点,该方法可以自动分割显微镜载玻片图像中的单个核和AgNOR,并计算每个核内AgNOR的数量。
方法:我们系统地定义了,在寻找高性能解决方案的过程中训练和测试了102个CNN。这包括对51个网络体系结构的评估,该网络体系结构将17个编码器与3个解码器和2个损耗功能结合在一起。这些CNN在一个新的AgNOR染色的口腔粘膜上皮细胞图像数据集上进行了训练和评估,其中包含来自48名患者的1,171张图像。有专家注释的地面真相。在我们的项目中开发的半自动程序极大地促进了注释。重叠的核,倾向于隐藏AgNOR,从而影响他们的真实计数,使用我们项目中也开发的自动解决方案丢弃。除了对测试数据集的评估,针对一组人类专家在第二个数据集上产生的结果,评估了最佳性能模型的稳健性.
结果:测试数据集上表现最好的CNN模型由具有焦点损失的DenseNet-169+LinkNet组成(DenseNet-169作为编码器,LinkNet作为解码器)。它获得了0.90的骰子得分和0.84的交集(IoU)。原子核和AgNORs的计数实现了原子核的准确率和召回率分别为0.94和0.90,AgNORs为0.82和0.74,分别。我们的解决方案在来自6名新患者的291张图像上实现了与人类专家相似的性能,获得核内相关系数(ICC)为0.91,AgNORs为0.81,95%置信区间为[0.89,0.93]和[0.77,0.84],分别,和p值<0.001,证实其统计学意义。就患者数量而言,我们的AgNOR染色图像数据集是公开可用的最多样化的AgNOR染色图像数据集,并且是口腔细胞的第一个。
结论:基于CNN的联合分割和量化AgNOR染色图像中的细胞核和NOR达到了专家般的性能水平,同时比后者快几个数量级。我们的解决方案通过与一组专家产生的结果表现出强烈的一致性来证明这一点,强调其加速诊断工作流程的潜力。我们训练有素的模型,代码,和数据集可用,可以刺激早期口腔癌检测的新研究。
OBJECTIVE: Oral cancer is the sixth most common kind of human cancer. Brush cytology for counting Argyrophilic Nucleolar Organizer Regions (AgNORs) can help early mouth cancer detection, lowering patient mortality. However, the manual counting of AgNORs still in use today is time-consuming, labor-intensive, and error-prone. The goal of our work is to address these shortcomings by proposing a convolutional neural network (CNN) based method to automatically segment individual nuclei and AgNORs in microscope slide images and count the number of AgNORs within each nucleus.
METHODS: We systematically defined, trained and tested 102 CNNs in the search for a high-performing solution. This included the evaluation of 51 network architectures combining 17 encoders with 3 decoders and 2 loss functions. These CNNs were trained and evaluated on a new AgNOR-stained image dataset of epithelial cells from oral mucosa containing 1,171 images from 48 patients, with ground truth annotated by specialists. The annotations were greatly facilitated by a semi-automatic procedure developed in our project. Overlapping nuclei, which tend to hide AgNORs, thus affecting their true count, were discarded using an automatic solution also developed in our project. Besides the evaluation on the test dataset, the robustness of the best performing model was evaluated against the results produced by a group of human experts on a second dataset.
RESULTS: The best performing CNN model on the test dataset consisted of a DenseNet-169 + LinkNet with Focal Loss (DenseNet-169 as encoder and LinkNet as decoder). It obtained a Dice score of 0.90 and intersection over union (IoU) of 0.84. The counting of nuclei and AgNORs achieved precision and recall of 0.94 and 0.90 for nuclei, and 0.82 and 0.74 for AgNORs, respectively. Our solution achieved a performance similar to human experts on a set of 291 images from 6 new patients, obtaining Intraclass Correlation Coefficient (ICC) of 0.91 for nuclei and 0.81 for AgNORs with 95% confidence intervals of [0.89, 0.93] and [0.77, 0.84], respectively, and p-values < 0.001, confirming its statistical significance. Our AgNOR-stained image dataset is the most diverse publicly available AgNOR-stained image dataset in terms of number of patients and the first for oral cells.
CONCLUSIONS: CNN-based joint segmentation and quantification of nuclei and NORs in AgNOR-stained images achieves expert-like performance levels, while being orders of magnitude faster than the later. Our solution demonstrated this by showing strong agreement with the results produced by a group of specialists, highlighting its potential to accelerate diagnostic workflows. Our trained model, code, and dataset are available and can stimulate new research in early oral cancer detection.