AICellCounter: A Machine Learning-Based Automated Cell Counting Tool Requiring Only One Image for Training
AICellCounter: A Machine Learning-Based Automated Cell Counting Tool Requiring Only One Image for Training
复制标题
AICellCounter：一种基于机器学习的自动细胞计数工具，仅需要一张图像进行训练
DOI：
10.1007/s12264-022-00895-w
复制
发表时间：
2022
影响因子：
5.6
通讯作者：
T. Gao
中科院分区：
文献类型：
--
作者：
Junnan Xu;Andong Wang;Yunfeng Wang;Jingting Li;Ruxia Xu;Hao Shi;Xiao;Yucheng Liang;Jian;T. Gao
高级特征（例如，相邻小区之间的关系），以实现更好的性能。许多基于AI的细胞分割研究遵循著名的R-CNN（区域卷积神经网络），Fast R-CNN和Mask R-CNN [1-3]架构。某些分割模型结合了联合收割机深度学习方法（例如，U-Net [4]）和传统方法（分水岭算法[5]）。此外，还有一些其他方法，包括GAN（生成对抗网络）[6]，StarDist [7]和TensorMask [8]。尽管人们已经开发了许多测试数据准确度很高的模型，但我们发现纸面上的算法与实践中可用的工具之间仍然存在相当大的差距。首先，在大量的标记的目标图像，这是不可行的最终用户由于高的标记成本。其次，尽管已经提出了许多少拍或零拍细胞分割模型，但它们仍然需要高性能的GPU（图形处理单元）嵌入式计算机进行训练，复杂的超参数设置和较长的微调时间。最后，许多方法只提供源代码，而不是用户友好的软件。我们开发了一个简单有效的自动化工具AICellCounter，通过结合深度学习（用于特征提取）和传统机器学习方法（用于结构简单和对新数据的鲁棒性）来进行细胞检测和计数。据我们所知，AICellCounter是基于AI的细胞计数工具的第一次尝试，（1）快速（根据样本图像的大小，训练时间在几秒到几分钟内），（2）只需要一个带有少量标记单元的样本图像，（3）即使在没有GPU的普通计算机上也能很好地执行，以及（4）具有简单的结构和很少的手动设置参数，这使得用户的学习曲线最小化。实验结果证明了AICellCounter的可靠性和灵活性。软件和用户视频亲爱的编辑，
high-level features (e.g., the relationships between neighboring cells) from the training data to achieve better performance. Many AI-based cell segmentation studies followed the famous R-CNN (region convolutional neural networks), Fast R-CNN, and Mask R-CNN [1–3] architectures. Some segmentation models combine deep learning approaches (e.g., U-Net [4]) and traditional approaches (Watershed algorithm [5]). In addition, there are some other approaches that include GAN (generative adversarial networks) [6], StarDist [7], and TensorMask [8]. Although numerous models with high accuracy on testing data have been developed, we find that a considerable gap still exists between the algorithms on paper and usable tools in practice. First, on lots of labeled target images, which is not feasible for end-users due to the high labeling cost. Second, even though many few-shot or zero-shot cell segmentation models have been proposed, they still require high-performance GPU (graphics processing unit)-embedded computers for training, complex hyper-parameter settings, and long fine-tuning times. Last, many methods only provide source code, not user-friendly software. we develop a simple and effective automated tool named AICellCounter to conduct cell detection and counting by combining deep learning (for feature extraction) and conventional machine learning method (for simple structure and robustness to new data). To our best knowledge, AICellCounter presents the first attempt of AI-based cell counting tool which (1) is fast (the training time is within a few seconds to a few minutes depending on the size of the sample image), (2) only needs one sample image with a few labeled cells, (3) performs well even on normal computers without GPUs, and (4) has a simple structure and few manually set parameters that minimize the learning curve for users. Experiment results demonstrated the reliability and flexibility of AICellCounter. The software and user video Dear Editor,
影响因子：
48
作者：
Falk, Thorsten;Mai, Dominic;Ronneberger, Olaf
通讯作者：
Ronneberger, Olaf