Methodology for Creating Annotated Datasets of Biological Objects in Microscopic Images

The study presents a methodology for automatically creating an annotated dataset for further use in training neural networks. The method is based on the Crisp-DM framework and consists of 6 stages: business understanding, data understanding, data preparation, modeling, evaluation, deployment, which has been modified due to the specifics of computer vision tasks. As a modeling method, the k-means method is proposed, which is applied to the normalized area and average color data of a given object. The approach is tested for creating an annotated dataset for recognizing and localizing cells of different types in blood smear images. The output is a dataframe with the columns file_name, xmin, ymin, xmax, ymax, type, area, which represent the file name, the coordinates of the upper left and lower right points of the bounding boxes, the cell type, and its area. Using these parameters and image processing libraries it is easy to add bounding boxes to the original images. The resulting annotated dataset is analyzed. Metrics were calculated and a loss matrix was constructed. All calculations were performed using the Python programming language and the Pandas, numpy, openCV, and matplotlib libraries.