Thin blood smear images of red blood cells with rouleaux formation morphology and normal morphology

This dataset contains images of thin blood smear with normal red blood cell morphology and rouleaux red blood cell morphology. Ethical approval with approval number: NHREC/17/03//2018 was obtained from Kano state ministry of health. Blood samples from 100 malaria infected patients were collected from Asiya Bayero pediatric hospital, kano state, Nigeria. Thick and thin blood smear slides were prepared using field stain. To ensure there was no bias in slide preparation, slides used for hospital diagnosis prepared under limited and constrained conditions were used as such types of slides represent the true reality of malaria diagnosis in less developed countries.Thin blood smear microscopy was performed by an expert microscopist and each slide was labeled according to the presence of Rouleaux formation or not among others. Out of 100 samples collected, 28 samples had rouleaux formation morphology. A 12MP iPhone 10 camera was attached to a microscope’s eyepiece. Pictures of different field of views for each slide were captured using the iPhone’s camera. For each slide, a minimum of 10 different field of views were captured. 616 images were captured for slides with rouleaux formation. To create a balanced dataset an equal number, 616 images were also captured for slides with normal morphology. To increase the size and variation of the dataset. 312 Digital images of thin blood smear slides with Giemsa staining collected from Murtala Muhammad specialist hospital were added. out of the 312 images, 156 had rouleaux RBC morphology and 156 had normal RBC morphology. Image capture was conducted in the morning, afternoon and evening and in different rooms with different lighting conditions to introduce diverse levels of illumination in the images The captured images from both hospitals had a size of 4032x3024 pixels. The background of the images were cropped to give a size 2500x2500 which were then sliced to give a final size of 750x750 pixels. The final data set consists of 12,356 thin blood smear images with rouleaux formation morphology and 12,356 thin blood smear images with normal red blood cell morphology. Different CNN architectures were trained for the binary classification of the dataset.本数据集收录了形态正常的红细胞与缗钱状红细胞（rouleaux red blood cell）的薄血涂片图像。本研究已获得尼日利亚卡诺州卫生部的伦理批准，批准编号为NHREC/17/03//2018。研究团队从尼日利亚卡诺州阿西娅·巴耶罗儿童医院采集了100份疟疾病患者的血液样本。采用现场染色法制备厚、薄血涂片玻片。为避免玻片制备过程引入偏倚，本研究选用了在有限约束条件下完成医院常规诊断的玻片——此类玻片能够反映欠发达国家疟疾诊断的真实场景。由专业显微技师开展薄血涂片镜检，并根据是否存在缗钱状红细胞聚集（rouleaux formation）等特征对每张玻片进行标注。在采集的100份样本中，有28份样本呈现缗钱状红细胞形态。 将1200万像素的iPhone 10摄像头安装至显微镜目镜，通过该iPhone摄像头采集每张玻片的不同视野图像。每张玻片至少采集10个不同视野的图像。其中，呈现缗钱状红细胞形态的玻片共采集到616张图像。为构建平衡数据集，研究团队为形态正常的红细胞玻片也采集了等量的616张图像。为扩充数据集规模与丰富数据多样性，本研究还纳入了从穆尔塔拉·穆罕默德专科医院采集的312张吉氏染色（Giemsa staining）薄血涂片数字图像：其中156张带有缗钱状红细胞形态，156张为正常红细胞形态。图像采集覆盖早、中、晚三个时段，并在不同房间、不同光照条件下完成，以引入多样化的光照水平。两所医院采集的原始图像分辨率均为4032×3024像素，研究人员对图像背景进行裁剪，将其调整为2500×2500像素，随后进一步切片得到最终尺寸为750×750像素的图像。本数据集最终包含12356张缗钱状红细胞形态的薄血涂片图像，以及12356张正常红细胞形态的薄血涂片图像。针对该数据集的二分类任务，研究团队已对多种卷积神经网络（CNN，Convolutional Neural Network）架构开展训练。