LeukocyteCount: Automatic Identification and Counting for leukocytes using Deep Learning

Manual white blood cell (WBC) counting is a labor-intensive and error-prone task in clinical hematology. This study proposes an automated, machine learning-based pipeline to improve the efficiency and accuracy of identifying, counting, and classifying leukocytes. By combining object detection with deep learning-based feature extraction, the authors aim to provide a robust tool for diagnosing immune-related blood conditions.

The proposed model operates in two primary stages:

1. Segmentation (YOLOv5): The researchers utilized the YOLOv5 architecture to detect and localize three types of blood cells: white blood cells, red blood cells, and platelets. This stage ensures that individual WBCs are isolated and centered for subsequent analysis.
2. Classification (MobileNetV2 + Logistic Regression): Once isolated, the WBC images are processed by a pre-trained MobileNetV2 model, which acts as a feature extractor. These extracted features are then fed into a Logistic Regression classifier to categorize the cells into four distinct types: Eosinophil, Lymphocyte, Monocyte, and Neutrophil.

The model was trained and validated using the publicly available Blood Cell Count Dataset (BCCD), leveraging transfer learning to adapt the pre-trained MobileNetV2 weights to the specific characteristics of blood smear images.

The hybrid approach demonstrated high performance across all metrics. The YOLOv5 segmentation module achieved 99.14% accuracy in detecting blood cells, while the classification module reached an accuracy of 98.14%. The authors note that the integration of MobileNetV2 with Logistic Regression provides a computationally efficient alternative to training deep networks from scratch, with an average pipeline prediction time of 1.5 seconds per image. The study also reports a significant improvement in F1 scores for red blood cell detection compared to previous benchmarks.