Deep Learning Applications in Hematology: Optimizing Accuracy and Robustness of Blood Cell Counting and Classification

Processing time, variability, and accuracy are the most significant limitations of traditional methods to determine the type and count of blood cells. These limitations could be addressed by using machine learning algorithms. Advanced YOLO architectures are among the most prominent of these technologies. This study aims to compare the performance of YOLOv11 and YOLOv12 models. Additionally, the impact of attention mechanisms on detection performance and the reduction of errors caused by cell overlap and interference were analyzed. Therefore, YOLOv11 and YOLOv12 models were applied to the BCCD dataset to detect, classify, and count blood cells (RBCs, WBCs, and platelets). Data augmentation strategies were implemented to increase the training dataset size. Both models were trained using a learning rate of 0.001, a batch size of 16, and 100 training epochs. The results showed that the YOLOv12 model achieved slightly higher numerical performance than YOLOv11 by achieving a validation mAP@50 value of 95.3% and a validation recall rate of 95.2%. The statistical results indicated that these differences were not statistically significant (p = 0.312). Therefore, both models provide comparable performance on the BCCD dataset. These findings suggest that both models are suitable for automated blood cell detection. However, larger and more diverse datasets are required to determine whether the observed numerical improvements generalize to broader clinical settings.