Clinical Dataset Of Bronchoalveolar Lavage Fluid Cell

This dataset comprises BALF cell images stained by Diff-quick staining and Gram staining, from patients who underwent bronchoalveolar lavage and endotracheal aspirates between 2018 and 2024 at the Chinese PLA General Hospital. The dataset contained 2105 images, with 13,263 annotated cells from seven typical cell classes, including erythrocyte, ciliated columnar epithelial, squamous epithelial, macrophage, lymphocyte, neutrophil, and eosinophil cells using both contour fine labeling and bounding box labeling. All cells that could be confidently identified by senior clinical cytologists were annotated. A small number of cells, for which consensus could not be reached among the cytologists, were left unannotated due to uncertainty. In addition, the classical YoloV8 deep learning instance segmentation model was used to detect and segment the seven types of BALF cells. We uploaded the best YOLO model weights as a supplement. As observed, our model demonstrated exceptionally high accuracy. There are four types of the dataset: High Resolution Images, Images, Visualization Images, and Labels, which are helpful to promote the study of automated cell identification in BALF. High Resolution Images: a total of 1903 original images (4,912×3,684 pixels) from clinical samples collected by an ImageView optical microscope (202 images were missed because of the initial anonymous input, but were all included in Images). Images: 2105 images from High Resolution Images resampled to the size of 853×640 for bounding box annotation and pixel-level high quality annotation, then applied directedly for network training. Visualization Images: images that showed manually labeled contours on each cell. Labels: saved in a txt format for direct use of YOLO series models and a convenient way to convert to other data formats.本数据集包含经Diff-quick染色法（Diff-quick staining）与Gram染色法（Gram staining）染色的支气管肺泡灌洗液（bronchoalveolar lavage fluid, BALF）细胞图像，采集自2018至2024年间在中国人民解放军总医院（Chinese PLA General Hospital）接受支气管肺泡灌洗及气管吸引术的患者。 数据集共计包含2105张图像，涵盖7类典型细胞的13263个标注细胞，包括红细胞（erythrocyte）、纤毛柱状上皮细胞（ciliated columnar epithelial cell）、鳞状上皮细胞（squamous epithelial cell）、巨噬细胞（macrophage）、淋巴细胞（lymphocyte）、中性粒细胞（neutrophil）及嗜酸性粒细胞（eosinophil cell），标注方式同时采用精细化轮廓标注（contour fine labeling）与边界框标注（bounding box labeling）。所有经资深临床细胞学家可明确鉴别的细胞均已完成标注；少量存在标注分歧、无法达成共识的细胞因存在不确定性未予以标注。 此外，本研究采用经典的YOLOv8深度学习实例分割模型对7类BALF细胞进行检测与分割，并上传了最优YOLO模型权重作为补充。经测试，该模型展现出极高的检测精度。本数据集包含四类文件：高分辨率图像（High Resolution Images）、普通图像（Images）、可视化图像（Visualization Images）与标注文件（Labels），可为BALF自动化细胞识别相关研究提供有力支撑。 高分辨率图像（High Resolution Images）： 共计1903张原始图像（分辨率为4912×3684像素），由ImageView光学显微镜（ImageView optical microscope）采集自临床样本（因初始匿名输入问题缺失202张，但该202张均已包含在普通图像集中）。 普通图像（Images）： 将高分辨率图像重采样至853×640分辨率，用于边界框标注与像素级高精度标注，后直接用于网络训练。 可视化图像（Visualization Images）： 带有手动标注细胞轮廓的图像。 标注文件（Labels）： 以TXT格式（txt format）存储，可直接用于YOLO系列模型，同时便于转换为其他数据格式。