A Comparative Study of YOLO Architectures with AI-Based Data Augmentation for Automated Platelet Estimation in Thrombocytopenic Patients

Abstract

Thrombocytopenia, a hematological disorder characterized by an abnormally low platelet count, demands diagnostic methods that are both rapid and precise. Conventional manual estimation of platelets from blood smears is notoriously labor-intensive and suffers from significant inter-observer variability. This study presents a systematic comparison of various YOLO object detection architectures, enhanced by advanced AI-based data augmentation, for the automated quantification of platelets in thrombocytopenic patients. We evaluated YOLOv5, YOLOv7, and YOLOv8 on a dataset of 1,500 annotated blood smear images from patients with confirmed low platelet counts. To overcome data scarcity, we implemented sophisticated augmentation techniques, including Generative Adversarial Networks (GANs) and neural style transfer. Our results establish the superior performance of YOLOv8, particularly when trained on GAN-generated samples, achieving a mean Average Precision ([email protected]) of 96.2%, a precision of 97.1%, and a recall of 95.8%. Crucially, the model sustained high accuracy across varying platelet densities and exhibited an almost perfect correlation (r = 0.98) with manual expert counts. This methodology lays the groundwork for a reliable, automated platelet assessment tool with the potential to significantly advance diagnostic and monitoring protocols for thrombocytopenia in clinical practice.