MICROSCOPIC BLOOD CELL SEGMENTATION USING DETR VISION TRANSFORMER

Abstract

ABSTRACT Manual microscopic blood cell analysis, a cornerstone of disease diagnosis, is plagued by inefficiency, subjectivity, and high costs, particularly in resource-limited settings like Sri Lanka. Existing automated hematology analyzers offer an alternative but remain largely inaccessible. This research introduces BCells-DETR, a novel model based on the DEtection TRansformer (DETR) architecture, designed for automated blood cell segmentation and classification from microscopic images. The model was trained using the Peripheral Blood Cells (PBC) dataset, compressing 5000 labeled images split 80/10/10 for training, validation, and testing. BCells- DETR performs precise instance segmentation and classification of erythrocytes (red blood cells) and leukocyte subtypes (e.g., neutrophils, eosinophils, basophils, monocytes, and lymphocytes). For real-world deployment, the trained model was integrated into a web application featuring a React frontend and Flask backend, allowing users to upload microscopic images and receive segmentation maps and cell counts in under a few seconds. On standard Common Objects in Context (COCO) evaluation metrics, the model achieved a mean Average Precision (mAP) of 0.800 and an Average Recall (AR) of 0.860, demonstrating high reliability. By providing rapid, quantitative results, this system mitigates the labor intensity and interobserver variability of manual methods. By successfully deploying a sophisticated deep learning model in an accessible platform, BCells-DETR presents a viable pathway toward democratizing computational pathology and augmenting diagnostic capabilities in low-resource environments. Keywords: Blood cell analysis, DETR, Vision Transformers, Image segmentation, Image classification, Biomedical imaging, Machine Learning, Web apps technology, PBC dataset