IEEE J Biomed Health Inform. 2022 Nov;26(11):5575-5583. doi: 10.1109/JBHI.2022.3203893. Epub 2022 Nov 11.
Precise and quick monitoring of key cytometric features such as cell count, size, morphology, and DNA content is crucial in life science applications. Traditionally, image cytometry relies on visual inspection of hemocytometers. This approach is error-prone due to operator subjectivity. Recently, deep learning approaches have emerged as powerful tools enabling quick and accurate image cytometry applicable to different cell types. Leading to simpler, compact, and affordable solutions, these approaches revealed image cytometry as a viable alternative to flow cytometry or Coulter counting. In this study, we demonstrate a modular deep learning system, DeepCAN, providing a complete solution for automated cell counting and viability analysis. DeepCAN employs three different neural network blocks called Parallel Segmenter, Cluster CNN, and Viability CNN that are trained for initial segmentation, cluster separation, and viability analysis. Parallel Segmenter and Cluster CNN blocks achieve accurate segmentation of individual cells while Viability CNN block performs viability classification. A modified U-Net network, a well-known deep neural network model for bioimage analysis, is used in Parallel Segmenter while LeNet-5 architecture and its modified version Opto-Net are used for Cluster CNN and Viability CNN, respectively. We train the Parallel Segmenter using 15 images of A2780 cells and 5 images of yeasts cells, containing, in total, 14742 individual cell images. Similarly, 6101 and 5900 A2780 cell images are employed for training Cluster CNN and Viability CNN models, respectively. 2514 individual A2780 cell images are used to test the overall segmentation performance of Parallel Segmenter combined with Cluster CNN, revealing high Precision/Recall/F1-Score values of 96.52%/96.45%/98.06%, respectively. Cell counting/viability performance of DeepCAN is tested with A2780 (2514 cells), A549 (601 cells), Colo (356 cells), and MDA-MB-231 (887 cells) cell images revealing high analysis accuracies of 96.76%/99.02%, 93.82%/95.93%, and 92.18%/97.90%, 85.32%/97.40%, respectively.
精确和快速监测关键细胞计量特征，如细胞计数、大小、形态和 DNA 含量，在生命科学应用中至关重要。传统上，图像细胞计量依赖于血球计数器的目视检查。由于操作人员的主观性，这种方法容易出错。最近，深度学习方法已经成为一种强大的工具，可以实现快速准确的适用于不同细胞类型的图像细胞计量。这些方法导致了更简单、更紧凑和更经济实惠的解决方案，它们将图像细胞计量揭示为流式细胞术或库尔特计数的可行替代方案。在这项研究中，我们展示了一个模块化的深度学习系统 DeepCAN，为自动细胞计数和活力分析提供了完整的解决方案。DeepCAN 采用了三个不同的神经网络块，称为 Parallel Segmenter、Cluster CNN 和 Viability CNN，它们分别用于初始分割、聚类分离和活力分析。Parallel Segmenter 和 Cluster CNN 块实现了单个细胞的精确分割，而 Viability CNN 块则进行了活力分类。我们在 Parallel Segmenter 中使用了一种著名的生物图像分析深度学习模型——修改后的 U-Net 网络，而在 Cluster CNN 和 Viability CNN 中分别使用了 LeNet-5 架构及其修改后的 Opto-Net。我们使用 15 张 A2780 细胞图像和 5 张酵母细胞图像来训练 Parallel Segmenter，总共包含 14742 张单个细胞图像。同样，我们使用 6101 张和 5900 张 A2780 细胞图像分别训练 Cluster CNN 和 Viability CNN 模型。我们使用 2514 张单独的 A2780 细胞图像来测试 Parallel Segmenter 与 Cluster CNN 相结合的整体分割性能，分别显示出 96.52%/96.45%/98.06%的高精度/召回率/F1-得分值。DeepCAN 的细胞计数/活力性能通过 A2780（2514 个细胞）、A549（601 个细胞）、Colo（356 个细胞）和 MDA-MB-231（887 个细胞）细胞图像进行了测试，分别显示出 96.76%/99.02%、93.82%/95.93%和 92.18%/97.90%、85.32%/97.40%的高分析准确性。
