Deep Learning Pipeline for Blood Cell Segmentation, Classification and Counting

Blood cell analysis, an important component of modern medicine, provides critical insights into a range of health conditions, right from minor infections to blood disorders such as leukemia. Traditional methods such as flow cytometry have been very effective in blood cell counting but are often reliant on expensive instruments, specialised reagents, and expert operators, thereby limiting their accessibility. To combat these challenges, this research explores the use of advanced deep learning methods to develop a robust and scalable pipeline for blood cell segmentation, classification, and counting. The proposed pipeline automates the entire process of generating high-precision segmentation masks and classifying blood cells into red blood cells or white blood cells by leveraging Meta AI’s Segment Anything Model and a Convolutional Neural Network. Apart from employing a host of image processing operations, the study also addresses challenges such as low image quality and data scarcity, which leads to class imbalance, by employing techniques such as image contrast enhancement and data augmentation. The pipeline can provide accurate cell counting from simple blood smears. The pipeline was then adapted to blood smear images collected by our research team through different instrumentation using transfer learning applied to the deep learning classifier. With over 95% accuracy in blood cell counts across different datasets, the pipeline exhibits a high degree of adaptability, which ultimately validates its potential for broader clinical applications.