Machine Learning-Based Normal White Blood Cell Multi-Classification Optimization

Le résumé fourni par la source

Clinically, the proportion and classification of white blood cell (WBC)s are currently established using manual methods, which rely on subjective judgment. Therefore, many studies are being conducted to automate the classification of WBC types. Several studies have employed deep learning (DL) or machine learning (ML) methods; no significant difference in performance between the two methods has been demonstrated. However, if the feature extraction and selection processes are optimized when using ML for WBC classification, its performance is improved. Therefore, in this study, we proposed an ML-based optimization system for five normal WBC classifications. Open datasets, Raabin-WBC, and private data were used. WBCs were segmented into nucleus, cytoplasm, and cell regions using U-Net, a DL model. The nucleus showed high segmentation performance with an average accuracy of 98.58% and a Dice coefficient of 0.9233, whereas the cells achieved an average accuracy of 99.47% and a Dice coefficient of 0.9324. Among the five multiple classifiers, the support vector machine achieved the highest accuracy of 97.36% and was chosen for WBC classification. Features used for WBC classification were determined through three experiments, resulting in a final selection of 108 features combining intensity histogram, hue saturation value, and CIE La$^{\ast }$b$^{\ast }$features. When using the selected features, the classification accuracy was 98.22%, indicating a high performance. Segmentation and classification of WBCs were performed using a graphical user interface and required approximately 137 s. The proposed system in this study is expected to enhance the efficiency of the existing PBS tests.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.