DeepBacs – Escherichia coli growth stage object detection dataset and YOLOv2 model

Training and test images of E. coli cells for object detection and classification using YOLOv2, as well as a trained YOLOv2 model. Additional information can be found on this github wiki. The example shows a bright field image of live <em>E. coli</em> cells and the respective annotation for specific growth stages. <strong>Training and test dataset</strong> <strong>Data type</strong>: Paired microscopy images (bright field) and annotations in PASCAL VOC format <strong>Microscopy data type</strong>: 2D bright field images recorded at 1 min interval <strong>Microscope</strong>: Nikon Eclipse Ti-E equipped with an Apo TIRF 1.49NA 100x oil immersion objective <strong>Cell type</strong>: <em>E. coli</em> MG1655 wild type strain (CGSC #6300). <strong>File format</strong>: .png (8-bit) <strong>Image size</strong>: 256 x 256 px² (158 nm / pixel), 100/15 individual frames (training/test dataset) 1024 x 1024 px² (79 nm / pixel), 9 regions of interest with 80 frames @ 1 min time interval (live-cell time series) <strong>Image preprocessing</strong>: Raw images were recorded in 16-bit mode (image size 512x512 px² @ 158 nm/px). 256 x 256 px² patches were extracted from individual frames and converted into 8-bit .png images after adjusting brightness and contrast. Annotation was performed online using <em>LabelImg </em>(https://github.com/tzutalin/labelImg). <strong>YOLOv2 model</strong> The YOLOv2 model was generated using the ZeroCostDL4Mic platform (Chamier et al., 2021). It was trained from scratch for 97 epochs on 100 manually annotated images (image dimensions: (256, 256)) with a batch size of 8 and a custom loss function combining MSE and crossentropy losses, using the YOLOv2 ZeroCostDL4Mic notebook (v 1.12.1). Key python packages used include tensorflow (v 0.1.12), Keras (v 2.3.1), numpy (v 1.19.5), cuda (v 11.0.221). The training was accelerated using a Tesla T4 GPU and data were augmented by a factor of 4 using flipping and rotation. The model weights can be used with the ZeroCostDL4Mic YOLOv2 notebook. <strong>Author(s)</strong>: Christoph Spahn^1,2, Mike Heilemann^1,3 <strong>Contact email</strong>: [e-mail] <strong>Affiliation(s)</strong>: 1) Institute of Physical and Theoretical Chemistry, Max-von-Laue Str. 7, Goethe-University Frankfurt, 60439 Frankfurt, Germany 2) ORCID: 0000-0001-9886-2263 3) ORCID: 0000-0002-9821-3578本数据集包含用于基于YOLOv2开展目标检测与分类的<em>大肠杆菌（E. coli）</em>细胞训练及测试图像，同时附带已训练完成的YOLOv2模型。更多详细信息可参阅该GitHub维基页面。本示例展示了活<em>大肠杆菌（E. coli）</em>细胞的明场成像图，以及对应不同生长阶段的标注信息。<strong>训练与测试数据集</strong><strong>数据类型</strong>：配对的显微成像图（明场）与PASCAL VOC格式标注文件<strong>显微数据类型</strong>：以1分钟为间隔采集的二维明场图像<strong>显微镜型号</strong>：尼康Eclipse Ti-E显微镜，搭配Apo TIRF 1.49NA 100倍油浸物镜<strong>细胞类型</strong>：<em>大肠杆菌（E. coli）</em> MG1655野生型菌株（CGSC #6300）<strong>文件格式</strong>：8位PNG格式<strong>图像尺寸</strong>：训练/测试数据集采用256×256像素的图像，像素分辨率为158 nm/像素，共包含100张训练帧与15张测试帧；活细胞时间序列数据集则采用1024×1024像素的图像，像素分辨率为79 nm/像素，包含9个感兴趣区域，每个区域含80张以1分钟为间隔采集的帧。<strong>图像预处理</strong>：原始图像以16位模式采集（尺寸为512×512像素，分辨率158 nm/像素）。从单帧图像中裁剪出256×256像素的图像块，调整亮度与对比度后转换为8位PNG格式。标注工作通过在线工具<em>LabelImg（LabelImg）</em>完成，其开源地址为https://github.com/tzutalin/labelImg。<strong>YOLOv2模型</strong>：本YOLOv2模型基于ZeroCostDL4Mic平台构建（Chamier等，2021）。使用YOLOv2版ZeroCostDL4Mic笔记本（v1.12.1），从随机初始化状态开始训练：基于100张手动标注的图像（尺寸为(256,256)），批次大小设为8，采用结合均方误差（Mean Squared Error）与交叉熵损失的自定义损失函数，共训练97个轮次。所用核心Python库包括tensorflow（v0.1.12）、Keras（v2.3.1）、numpy（v1.19.5）与cuda（v11.0.221）。训练过程通过Tesla T4 GPU加速，采用翻转与旋转操作将数据增强4倍。模型权重可配合ZeroCostDL4Mic YOLOv2笔记本使用。<strong>作者</strong>：Christoph Spahn^1,2、Mike Heilemann^1,3<strong>联系邮箱</strong>：[e-mail]<strong>机构信息</strong>：1) 德国法兰克福歌德大学物理与理论化学研究所，马克斯·冯·劳厄大街7号，法兰克福60439；2) ORCID: 0000-0001-9886-2263；3) ORCID: 0000-0002-9821-3578