Using deep learning on microscopic images for white blood cell detection and segmentation to assist in leukemia diagnosis.

Early and accurate diagnosis of leukemia is essential to increase the chances of successful treatment. However, manual analysis of bone marrow cells is a time-consuming process prone to errors. With advancements in health technologies, automating the detection of leukemic cells has become indispensable to accelerating and improving diagnostic accuracy. This study proposes the Y-YOLOv10 model, which combines Principal Component Analysis (PCA) to reduce the dimensionality of images and the YOLOv10 model for real-time detection, distinguishing 12 types of white blood cells. To enhance the diversity and robustness of training, six public datasets were combined, resulting in a comprehensive and representative dataset. Additionally, seven advanced data augmentation techniques were applied, including rotation, zoom, blurring, noise addition, contrast adjustment, horizontal and vertical flips, and cropping. These techniques were used alongside the original data, generating realistic variations and improving the model’s ability to handle real-world laboratory scenarios. Furthermore, H.264 compression artifacts were introduced during training to simulate conditions of reduced image quality, making the Y-YOLOv10 robust to variations in acquisition quality. The Y-YOLOv10 achieved average performances of 96.85% accuracy, 95.21% recall, and 96.76% F1-score, outperforming methods such as YOLOv8 and ResNet50. Additionally, H.264 compression allowed for a 30–40% reduction in data size while maintaining high performance and faster processing times. This study highlights the potential of lightweight deep learning models to improve leukemia classification, reduce the workload of specialists, and ensure a robust and adaptable system for various application conditions. [ABSTRACT FROM AUTHOR]