AI-powered body fluid cell classification: Development and validation using Roboflow and YOLOv11n framework

Accurate classification of body fluid cells is crucial for diagnosing various medical conditions. Traditional manual methods are often time-consuming and subject to inter-observer variability. In this study, we developed and compared two AI-based classification models, YOLOv11n and Roboflow 3.0, for body fluid cell classification. A dataset comprising 12,966 cells across 14 distinct categories was used for training (70%), validation (20%), and testing (10%). Expert consensus from three medical technologists ensured robust labeling, achieving a Fleiss’ Kappa of 0.877. Both models were validated against expert consensus on unseen images, with YOLOv11n achieving a Kappa of 0.932 and Roboflow 3.0 a Kappa of 0.930, indicating almost perfect agreement. The Roboflow 3.0 model demonstrated a mean average precision (mAP50) of 78.9%, with a precision of 74.3%, recall of 75.0%, and F1 score of 74.6%, while YOLOv11n achieved an mAP50 of 73.9%, precision of 69.8%, recall of 75.0%, and F1 score of 72.3%. Both models showed high accuracy for common cell types but faced challenges with rarer categories and in distinguishing between morphologically similar cells like macrophages and monocytes. Notably, these models offer a consistent and efficient tool for cell classification, providing an advantage for educational purposes and facilitating the reskilling of medical technologists in the application of AI in diagnostics.