Efficient blood cell classification from microscopic smear images using U-Net segmentation and a lightweight CNN.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Manual blood cell classification is labor-intensive and error-prone, hindering timely diagnosis of diseases like leukemia and anemia.
• Automated systems are needed to improve the speed and accuracy of blood cell analysis.

Purpose of the Study:

• To develop and evaluate a deep learning-based automated system for classifying nine types of blood cells from microscopic images.
• To enhance the accuracy and efficiency of blood cell counting and classification for disease diagnosis.

Main Methods:

• Utilized U-Net for semantic segmentation of blood cells, followed by a watershed algorithm to separate overlapping cells.
• Developed a custom lightweight convolutional neural network (LWCNN) named BloodCell-Net for classifying individual segmented cells.
• Employed image preprocessing techniques including resizing, rescaling, contrast enhancement, and augmentation.

Main Results:

• The U-Net segmentation model achieved high performance with 98.23% accuracy, 98.40% precision, and 95.97% Intersection over Union (IOU).
• The BloodCell-Net classifier demonstrated excellent accuracy (97.10%), precision (97.19%), recall (97.01%), and F1 score (97.10%) for nine blood cell types.
• A 5-fold cross-validation was used to ensure model generalization and reduce overfitting.

Conclusions:

• The proposed deep learning system offers an accurate and efficient automated solution for blood cell classification and counting.
• This automated approach has the potential to significantly aid in the early diagnosis and management of various hematological disorders.
• The BloodCell-Net approach provides a robust method for analyzing microscopic blood smear images.