Deep Learning Detection and Classification of Red Blood Cells: Towards a Universal Dataset

We evaluate emerging machine learning models for pattern recognition, focusing on the YOLOv11 architecture for detecting and classifying red blood cell shapes. Our analysis targets two characteristic morphologies observed under flow: slipper and parachute. A key challenge in this task is the development of a robust and diverse dataset. To address this, we employ synthetic image generation using a cut-and-paste approach, introducing variations in cell overlap and arrangements of microfluidic channels disposition to alleviate data scarcity and reduce cross-dataset bias. We generate these datasets with U-Net and Cellpose segmentation models, and rigorously assess YOLOv11 performance on two benchmarks: (i) a controlled dataset for evaluating classification accuracy, and (ii) a challenging, visually heterogeneous dataset for assessing generalization. Results show that the model achieves high precision for distinct cell types in controlled settings, but exhibits reduced performance on the unseen dataset, highlighting a trade-off between specialized accuracy and broad applicability in complex microscopy scenarios.